Table of Content

10 October 2020, Volume 40 Issue 10

Previous Issue    Next Issue

Artificial intelligence

Hybrid recommendation algorithm based on rating filling and trust information

SHEN Xueli, LI Zijian, HE Chenhao

2020, 40(10):  2789-2794.  DOI: 10.11772/j.issn.1001-9081.2020020267

Asbtract ( )   PDF (904KB) ( )  

References | Related Articles | Metrics

Aiming at the problem of poor recommendation effect caused by the data sparsity of the recommendation system, a hybrid recommendation algorithm based on rating filling and trust information was proposed namely RTWSO (Real-value user item restricted Boltzmann machine Trust Weighted Slope One). Firstly, the improved restricted Boltzmann machine model was used to fill the rating matrix, so as to alleviate the sparseness problem of the rating matrix. Secondly, the trust and trusted relationships were extracted from the trust relationship, and the matrix decomposition based implicit trust relationship similarity was also used to solve the problem of trust relationship sparsity. The modification including trust information was performed to the original algorithm, improving the recommendation accuracy. Finally, the Weighted Slope One (WSO) algorithm was used to integrate the matrix filling and trust similarity information as well as predict the rating data. The performance of the proposed hybrid recommendation algorithm was verified on Epinions and Ciao datasets. It can be seen that the proposed hybrid recommendation algorithm has the recommendation accuracy improved by more than 3% compared with the composition algorithm, and recommendation accuracy increased by more than 1.2% compared with the existing social recommendation algorithm SocialIT (Social recommendation algorithm based on Implict similarity in Trust). Experimental results show that the proposed hybrid recommendation method based on rating filling and trust information, improves the recommended accuracy to a certain extent.

Recommendation method based on multidimensional social relationship embedded deep graph neural network

HE Haochen, ZHANG Danhong

2020, 40(10):  2795-2803.  DOI: 10.11772/j.issn.1001-9081.2020040569

Asbtract ( )   PDF (1544KB) ( )  

References | Related Articles | Metrics

The social recommendation system can alleviate the data sparsity and cold start problems in the recommendation system through the users' social attribute information, thereby improving the accuracy of the recommendation system. However, most social recommendation methods mainly aim at the single social network or linearly superimpose multiple social networks, making it difficult for the users' social attributes to fully participate in the calculation, so the accuracy of recommendation is limited. To solve this problem, a multi-network embedded graph neural network model was proposed to implement the recommendation in complex multidimensional social networks. In the model, a unified method was built to fuse the multidimensional complex networks composed of user-item, user-user and other relationships. Different types of multi-neighbors were aggregated to attribute to the node generation through attention mechanism, and multiple graph neural networks were combined to construct a graph neural network recommendation framework under multidimensional social relationships. In the proposed method, the entities in the recommendation system and their relationships were reflected by the topology structure, and the relevant information was calculated and updated continuously on the graph directly. It can be seen that the method is inductive, and avoids the problem of incomplete information utilization in traditional recommendation methods effectively. By comparing with related social recommendation algorithms, the experimental results show that the recommendation accuracy indicators such as Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) of the proposed method are improved, and the method even has good accuracy on sparse data.

Improved AdaNet based on adaptive learning rate optimization

LIU Ran, LIU Yu, GU Jinguang

2020, 40(10):  2804-2810.  DOI: 10.11772/j.issn.1001-9081.2020020237

Asbtract ( )   PDF (1134KB) ( )  

References | Related Articles | Metrics

AdaNet (Adaptive structural learning of artificial neural Networks) is a neural architecture search framework based on Boosting ensemble learning, which can create high-quality models through integrated subnets. The difference between subnets generated by the existing AdaNet is not significant, which limits the reduction of generalization error in ensemble learning. In the two steps of AdaNet:setting subnet network weights and integrating subnets, Adagrad, RMSProp (Root Mean Square Prop), Adam, RAdam (Rectified Adam) and other adaptive learning rate methods were used to improve the existing optimization algorithms in AdaNet. The improved optimization algorithms were able to provide different degrees of learning rate scaling for different dimensional parameters, resulting in a more dispersed weight distribution, so as to increase the diversity of subnets generated by AdaNet, thereby reducing the generalization error of ensemble learning. The experimental results show that on the three datasets:MNIST (Mixed National Institute of Standards and Technology database), Fashion-MNIST and Fashion-MNIST with Gaussian noise, the improved optimization algorithms can improve the search speed of AdaNet, and more diverse subnets generated by the method can improve the performance of the ensemble model. For the F1 value, which is an index to evaluate the model performance, compared with the original method, the improved methods have the largest improvement of 0.28%, 1.05% and 1.10% on the three datasets.

FPGA-based convolutional neural network fixed-point acceleration

LEI Xiaokang, YIN Zhigang, ZHAO Ruilian

2020, 40(10):  2811-2816.  DOI: 10.11772/j.issn.1001-9081.2020020256

Asbtract ( )   PDF (1063KB) ( )  

References | Related Articles | Metrics

Aiming at the problem of high running power consumption and slow operation of Convolutional Neural Network (CNN) on resource-constrained hardware devices, a method for accelerating fixed-point computation of CNN based on Field Programmable Gate Array (FPGA) was proposed. First, a fixed-point processing method was proposed. In order to reduce the storage space of the CNN parameters, different scale parameters were designed for different convolution layers and the relative divergence was used to determine the bit width length. The effect of different quantization intervals on the accuracy of CNN was studied. Then, the parameter multiplexing method and the pipeline calculation method were designed to accelerate the convolution calculation. In order to verify the acceleration effect of CNN after fixed-point processing, two datasets of face and ship were used for verification. Compared with the traditional floating-point convolution computation, on the premise of ensuring that the accuracy loss of the CNN is small, when the weight parameters and the input feature map parameters are quantized to 7-bit, on the face recognition CNN model, the proposed method has the compressed weight parameter file size of about 22% of the origin, and the convolution calculation speedup is 18.69. At the same time, the method makes the utilization rate of the multiplier-accumulator in FPGA reach 94.5%. Experimental results show that the proposed method can improve the speed of convolution calculation, and efficiently use FPGA hardware resources.

Bio-inspired matrix reduction and quantization method for deep neural network

ZHU Qianqian, LIU Yuan, LI Fu

2020, 40(10):  2817-2821.  DOI: 10.11772/j.issn.1001-9081.2020020222

Asbtract ( )   PDF (1067KB) ( )  

References | Related Articles | Metrics

Bio-inspired Deep Neural Network (DNN) is a revolutionary breakthrough in artificial intelligent field. However, the lack of storage space as well as computing capacity caused by the explosive increase of the model weights not only keeps DNN apart from its original inspiration, but also makes it difficult to deploy DNN on embedded/mobile devices. In order to solve this problem, the biological selection principle in the evolution was studied, and a novel neural network algorithm based on "evolution" + "randomness" + "selection" was proposed. In this method, the size of the existing models were greatly simplified on the premise of maintaining the basic framework of the existing neural network models. First, the weight parameters were clustered. Then, based on the cluster centroid values of the parameters, the random perturbation was added to reconstruct the parameters. Finally, the image classification and object detection were performed on the reconstructed model to realize the accuracy test and model stability analysis. Experimental results on ImageNet dataset and COCO dataset show that the proposed model reconstruction method can compress the sizes of four models, including Darknet19, ResNet18, ResNet50 and YOLOv3, to 1/4-1/3 of the original ones, and under the condition of 1%-3% performance improvement in the test accuracy of image classification and object detection, there is the possibility of further simplification.

Object tracking algorithm based on jointly-optimized strong-coupled Siamese region proposal network

SHI Guoqiang, ZHAO Xia

2020, 40(10):  2822-2830.  DOI: 10.11772/j.issn.1001-9081.2020030297

Asbtract ( )   PDF (2660KB) ( )  

References | Related Articles | Metrics

For the mismatch between the maximum classification score and the best border in Region Proposal Network (RPN) during the object tracking task, a object tracking algorithm based on Strong-Coupled Siamese Region Proposal Network (SCSiamRPN) was proposed. Firstly, the Bounded IoU method was adopted to optimize the calculation of Intersection over Union (IoU) value of positive samples, and the calculation process was simplified by decomposing formulas, fixing variables, substituting differences, and constraining approximation values. Then, the loss function structures were optimized. A coupling factor with the IoU value as bound was added to the classification loss function to combine the classification task and the border regression task, so as to increase the loss values of high IoU samples. A weighted coefficient with the IoU value as main variable was added to the border regression loss function to increase the contributions of the target center samples, so as to improve the border localization precision. Simulation results showed that the proposed algorithm had the tracking precision and success rate reached 0.86 and 0.64 respectively on OTB100 dataset, both of which were improved by 3% in the comparison with those of high performance visual tracking with Siamese Region Proposal Network (SiamRPN). It is found that the proposed algorithm solves the mismatch between the maximum classification score, enhances the coupling between the classification task and the border regression task, greatly improves the tracking precision without slowing down the tracking speed.

Distribution path planning and charging strategy for pure electric vehicles with load constraint

LIU Yuliang, CHEN Huaili

2020, 40(10):  2831-2837.  DOI: 10.11772/j.issn.1001-9081.2020020157

Asbtract ( )   PDF (899KB) ( )  

References | Related Articles | Metrics

Due to the limitation of driving mileage of pure electric vehicles, it is difficult to realize the long-distance transportation service of pure electric vehicles in a short time to meet the commercial requirements. However, due to the characteristics such as small distribution area, small quantity per batch and large batch number of urban logistics, the pure electric vehicles can be considered to complete the urban distribution tasks. In order to meet the requirements of multiple distribution tasks of the vehicle on the same day and consider the specific impact of vehicle load on real-time energy consumption, a distribution model considering the impact of vehicle load on real-time energy consumption was established to meet the customers' service time requirements in a timely manner. Taking city A as an example, an ant colony algorithm was designed to solve the model, so as to make the reasonable path planning and charging strategy arrangement for the distribution tasks of pure electric vehicles. Finally, the feasibility of pure electric vehicles in urban distribution and logistics in the future was analyzed by comparing to the operation with fuel vehicles.

Long text aspect-level sentiment analysis based on text filtering and improved BERT

WANG Kun, ZHENG Yi, FANG Shuya, LIU Shouyin

2020, 40(10):  2838-2844.  DOI: 10.11772/j.issn.1001-9081.2020020164

Asbtract ( )   PDF (1014KB) ( )  

References | Related Articles | Metrics

Aspect-level sentiment analysis aims to classify the sentiment of text in different aspects. In the aspect-level sentiment analysis of long text, the existing aspect-level sentiment analysis algorithms do not fully extract the features of aspect related information in the long text due to the redundancy and noise problems, leading to low classification accuracy. On the datasets with coarse and fine aspects, existing solutions do not take advantage of the information in the coarse aspect. In view of the above problems, an algorithm named TFN+BERT-Pair-ATT was proposed based on text filtering and improved Bidirectional Encoder Representation from Transformers (BERT). First, the Text Filter Network (TFN) based on Long Short-Term Memory (LSTM) neural network and attention mechanism was used to directly select part sentences related to the coarse aspect from the long text. Next, the related sentences were associated with others in order, and after combining with fine aspects, the sentences were input into the BERT-Pair-ATT, which is with the attention layer added to the BERT, for feature extraction. Finally, the sentiment classification was performed by using Softmax. Compared with the classical Convolutional Neural Network (CNN) based models such as Gated Convolutional network with Aspect Embedding (GCAE) and LSTM based model Interactive Attention Network (IAN), the proposed algorithm improves the related evaluation index by 3.66% and 4.59% respectively on the validation set, and improves the evaluation index by 0.58% compared with original BERT. Results show that the algorithm based on text filtering and improved BERT has great value in the aspect-level sentiment analysis task of long text.

Chinese-Vietnamese bilingual multi-document news opinion sentence recognition based on sentence association graph

WANG Jian, TANG Shan, HUANG Yuxin, YU Zhengtao

2020, 40(10):  2845-2849.  DOI: 10.11772/j.issn.1001-9081.2020020280

Asbtract ( )   PDF (815KB) ( )  

References | Related Articles | Metrics

The traditional opinion sentence recognition tasks mainly realize the classification by emotional features inside the sentence. In the task of cross-lingual multi-document opinion sentence recognition, the certain supporting function for opinion sentence recognition was provided by the association between sentences in different languages and documents. Therefore, a Chinese-Vietnamese bilingual multi-document news opinion sentence recognition method was proposed by combining Bi-directional Long Short Term Memory (Bi-LSTM) network framework and sentence association features. Firstly, emotional elements and event elements were extracted from the Chinese-Vietnamese bilingual sentences to construct the sentence association diagram, and the sentence association features were obtained by using TextRank algorithm. Secondly, the Chinese and Vietnamese news texts were encoded in the same semantic space based on the bilingual word embedding and Bi-LSTM. Finally, the opinion sentence recognition was realized by jointly considering the sentence coding features and semantic features. The theoretical analysis and simulation results show that integrating sentence association diagram can effectively improve the precision of multi-document opinion sentence recognition.

Automatic emotion annotation method of Yi language data based on double-layer features

HE Jun, ZHANG Caiqing, ZHANG Yunfei, ZHANG Dehai, LI Xiaozhen

2020, 40(10):  2850-2855.  DOI: 10.11772/j.issn.1001-9081.2020020148

Asbtract ( )   PDF (1335KB) ( )  

References | Related Articles | Metrics

Most of the existing automatic emotion annotation methods only extract the single recognition feature from acoustic layer or language layer. While Yi language is affected by the factors such as too many branch dialects and high complexity, so the accuracy of automatic annotation of Yi language with single-layer emotion feature is low. Based on the features such as rich emotional affixes in Yi language, a double-layer feature fusion method was proposed. In the method, the emotional features from acoustic layer and language layer were extracted respectively, the methods of generating sequence and adding units as needed were applied to complete the feature sequence alignment, and the process of automatic emotion annotation was realized through the corresponding feature fusion and automatic annotation algorithm. Taking the speech and text data of Yi language in a poverty alleviation log database as samples, three different classifiers were used for comparative experiments. The results show that the classifier has no obvious effect on the automatic annotation results, and the accuracy of automatic annotation after the fusion of double-layer features is significantly improved, the accuracy is increased from 48.1% of acoustic layer and 34.4% of language layer to 64.2% of double-layer fusion.

Multi-pose feature fusion generative adversarial network based face reconstruction method

LIN Leping, LI Sanfeng, OUYANG Ning

2020, 40(10):  2856-2862.  DOI: 10.11772/j.issn.1001-9081.2020020205

Asbtract ( )   PDF (3013KB) ( )  

References | Related Articles | Metrics

Concerning the problem that single face image is difficult to solve the large-pose profile face in face reconstruction, a face reconstruction method based on Multi-pose Feature Fusion Generative Adversarial Network (MFFGAN) was proposed. In this method, the relevant information between multiple profile faces with different poses was used for face reconstruction, and the adversarial mechanism was used to adjust network parameters. A new network was designed in the method, which consisted of a generator including multi-pose feature extraction, multi-pose feature fusion and frontal face synthesis, and a discriminator for adversarial training. In the multi-pose feature extraction module, multiple convolution layers were used to extract the multi-pose features of profile face images. In the multi-pose feature fusion module, the multi-pose features were fused into a fusion feature containing multi-pose face information. And, the fusion feature was added during the face reconstruction process in the frontal face synthesis module. Obtaining the relevant information and global structure by exploring the feature dependency between multi-pose profile face images can effectively improve the reconstruction results. Experimental results show that, compared with those of the state-of-the-art deep learning based face reconstruction methods, the contours of the frontal face recovered by the proposed method are clear, and the recognition rate of the frontal face recovered from two profile faces is increased by 1.9 percentage points on average; and the more profile faces are input, the higher the recognition rate of the recovered frontal face is, which indicates that the proposed method can effectively fuse multi-pose features to recover a clear frontal face.

Path planning algorithm of multi-population particle swarm manipulator based on monocular vision

YUAN Meng'en, CHEN Lijia, FENG Zikai

2020, 40(10):  2863-2871.  DOI: 10.11772/j.issn.1001-9081.2020020145

Asbtract ( )   PDF (1773KB) ( )  

References | Related Articles | Metrics

Aiming at the path planning problem of manipulator with complex static background and multiple constraints, a new multi-population particle swarm optimization algorithm based on elite population and monocular vision was proposed. Firstly, the image difference algorithm was used to eliminate the background, then the contour surrounding method was used to find out the target area, and the model pose estimation method was used to locate the target position. Secondly, a multi-population particle swarm optimization based on elite population was proposed to obtain the optimal angles of the manipulator according to the target position. In this algorithm, the elite population and the sub-populations were combined to form the multi-population particle swarm, and the pre-selection and interaction mechanisms were used to make the algorithm jump out of local optimums. The simulation results show that compared with the real coordinates, the coordinates error of the object position obtained by background elimination method is small; compared with those of the state-of-the-art evolutionary algorithms, the average fitness values of the paths and the Mean Square Errors (MSE) obtained by the proposed algorithm are the smallest for the objects in different positions.

Traffic sign recognition based on improved convolutional neural network with spatial pyramid pooling

DENG Tianmin, FANG Fang, ZHOU Zhenhao

2020, 40(10):  2872-2880.  DOI: 10.11772/j.issn.1001-9081.2020020214

Asbtract ( )   PDF (3595KB) ( )  

References | Related Articles | Metrics

In order to solve the problems of low accuracy and poor generalization of traffic sign recognition caused by factors such as fog, light, occlusion and large inclination, a lightweight traffic sign recognition method based on neural network was proposed. First, in order to improve image quality, the methods of image normalization, affine transformation and Contrast Limited Adaptive Histogram Equalization (CLAHE) were used for image preprocessing. Second, based on Convolutional Neural Network (CNN), spatial pyramid structure and Batch Normalization (BN) were fused to construct an improved CNN with Spatial Pyramid Pooling (SPP) and BN (SPPN-CNN), and Softmax classifier was used to perform the traffic sign recognition. Finally, the German Traffic Sign Recognition Benchmark (GTSRB) was used to compare the training effects of different image preprocessing methods, model parameters and model structures, and to verify and test the proposed model. Experimental results show that for SPPN-CNN model, the recognition accuracy reaches 98.04% and the loss is less than 0.1, and the recognition rate is greater than 3 000 frame/s under the condition of GPU with low configuration,verifying that the SPPN-CNN model has high accuracy, strong generalization and good real-time performance.

Vehicle classification based on HOG-C CapsNet in traffic surveillance scenarios

CHEN Lichao, ZHANG Lei, CAO Jianfang, ZHANG Rui

2020, 40(10):  2881-2889.  DOI: 10.11772/j.issn.1001-9081.2020020152

Asbtract ( )   PDF (3651KB) ( )  

References | Related Articles | Metrics

To improve the performance of vehicle classification by making full use of image information from traffic surveillance, Histogram of Oriented Gradient Convolutional (HOG-C) features extraction method was added on the capsule network, and a Capsule Network model fusing with HOG-C features (HOG-C CapsNet) was proposed. Firstly, the gradient data in the images were calculated by the gradient statistical feature extraction layer, and then the Histogram of Oriented Gradient (HOG) feature map was plotted. Secondly, the color information of the image was extracted by the convolutional layer, and then the HOG-C feature map was plotted with the extracted color information and HOG feature map. Finally, the HOG feature map was input into to the convolutional layer extract its abstract features, and the abstract features were encapsulated through a capsule network into capsules with the three-dimensional spatial feature representation, so as to realize the vehicle classification by dynamic routing algorithm. Compared with other related models on the BIT-Vehicle dataset, the proposed model has the accuracy of 98.17%, the Mean Average Precision (MAP) of 97.98%, the Mean Average Recall (MAR) of 98.42% and the comprehensive evaluation index of 98.20%. Experimental results show that the vehicle classification in traffic surveillance scenarios can be achieved with better performance by using HOG-C CapsNet.

Deep domain adaptation model with multi-scale residual attention for incipient fault detection of bearings

MAO Wentao, YANG Chao, LIU Yamin, TIAN Siyu

2020, 40(10):  2890-2898.  DOI: 10.11772/j.issn.1001-9081.2020030329

Asbtract ( )   PDF (2274KB) ( )  

References | Related Articles | Metrics

Aiming at the problems of poor reliability and high false alarm rate of the fault detection models of bearings caused by the differences in working environment and equipment status, a multi-scale attention deep domain adaptation model was proposed according to the characteristics and needs of incipient fault detection. First, the monitoring signal was pre-processed into a three-channel data consisting of the original signal, Hilbert-Huang transform marginal spectrum and frequency spectrum. Second, the filters of different sizes were added into the residual attention module to extract multi-scale deep features, and the convolution-deconvolution operation was used to reconstruct the input information in order to obtain attention information, then a multi-scale residual attention module was constructed by combining the attention information and multi-scale features and was used to extract the attention features with stronger ability of representing incipient faults. Third, a loss function based on the cross entropy and Maximum Mean Discrepancy (MMD) regularization constraints was constructed to achieve the domain adaptation on the basis of the extracted attention features. Finally, a stochastic gradient descent algorithm was used to optimize the network parameters, and an end-to-end incipient fault detection model was established. Comparative experiments were conducted on the IEEE PHM-2012 Data Challenge dataset. Experimental results show that, compared with eight representative incipient fault detection and diagnosis methods as well as transfer learning algorithms, the proposed method can obtain the reduction of 62.7% and 61.3% in the average false alarm rate while keeping the alarm location not delayed, and effectively improves the robustness of incipient fault detection.

Ampoule packaging quality inspection algorithm based on machine vision and lightweight neural network

GAO Ming, REN Dejun, HU Yunqi, FU Lei, QIU Lyu

2020, 40(10):  2899-2903.  DOI: 10.11772/j.issn.1001-9081.2020020143

Asbtract ( )   PDF (1784KB) ( )  

References | Related Articles | Metrics

Focusing on the problems such as low inspection speed and low accuracy caused by subjective factors in the manual inspection method of ampoule packaging quality, an inspection algorithm based on machine vision and lightweight neural network was proposed. First, threshold processing, tilt correction and cutting of ampoule regions were performed on the images to be inspected by using the threshold segmentation and affine transformation methods in machine learning. Second, the network structure of the classification algorithm was designed according to the characteristics of images and the requirements of defect recognition. Finally, the ampoule packaging defect dataset was constructed by collecting the images of the production site. After that, the proposed ampoule packaging defect identification network was verified, and the accuracy and inspection speed of the algorithm deployed on the Jetson Nano embedded platform were tested. Experimental results show that, taking the product of five ampoules each box as the example, the proposed ampoule packaging quality inspection algorithm takes 70.1 ms/box averagely, that is up to 14 boxes/s, and has the accuracy of 99.94%. It can achieve online high-precision ampoule packaging quality inspection on the Jetson Nano embedded platform.

Deep mixed convolution model for pulmonary nodule detection

QI Yongjun, GU Junhua, ZHANG Yajuan, WANG Feng, TIAN Zepei

2020, 40(10):  2904-2909.  DOI: 10.11772/j.issn.1001-9081.2020020192

Asbtract ( )   PDF (1572KB) ( )  

References | Related Articles | Metrics

Pulmonary nodule detection is a very challenging task based on high-dimensional lung Computed Tomography (CT) images. Among many pulmonary nodule detection algorithms, the deep Convolutional Neural Network (CNN) is the most attractive one. In this kind of networks, the Two-Dimensional (2D) CNNs with many pre-trained models and high detection efficiency are widely used. However, the nature of pulmonary nodules is the Three-Dimensional (3D) lesion, so that the 2D CNNs will inevitably cause information loss and thereby affect the detection accuracy. The 3D CNNs can make full use of the spatial information of CT images and effectively improve the detection accuracy, but the 3D CNNs have shortcomings such as many parameters, large calculation consumption and high risk of over fitting. In order to take the advantages of the two networks, a pulmonary nodule detection model based on a deep mixed CNN was proposed. By deploying 3D CNN in the shallow layer of the neural network model and 2D CNN in the deep layer of the model, and adding a deconvolution module to fuse multi-layer image features together, the model parameters were reduced and the generalization ability and the detection efficiency of the model were improved without decreasing the detection accuracy. Experimental results on LUNA16 dataset show that the proposed model has the sensitivity reached 0.924 under the condition of average 8 false positives per scan, which outperforms the existing state-of-the-art models.

Automatic segmentation of breast epithelial and stromal regions based on conditional generative adversarial network

ZHANG Zelin, XU Jun

2020, 40(10):  2910-2916.  DOI: 10.11772/j.issn.1001-9081.2020020162

Asbtract ( )   PDF (7615KB) ( )  

References | Related Articles | Metrics

The automatic segmentation of epithelial and stromal regions in breast pathological images has very important clinical significance for the diagnosis and treatment of breast cancer. However, due to the high complexity of epithelial and stromal regions in breast tissue pathological images, it is difficult for general segmentation models to effectively train the model based on the provided segmentation labels only, and perform fast and accurate segmentation of the two regions. Therefore, based on conditional Generative Adversarial Network (cGAN), the EPithelium and Stroma segmentation conditional Generative Adversarial Network (EPScGAN) model was proposed. In EPScGAN, the discrimination mechanism of the discriminator provided a trainable loss function for the training of the generator, in order to measure the error between the segmentation result outputs of the generator and the real labels more accurately, so as to better guide the generator training. Total of 1 286 images with the size of 512×512 were randomly cropped as an experimental dataset from the expert-labeled breast pathological image datasets provided by the Netherlands Cancer Institute (NKI) and the Vancouver General Hospital (VGH). Then the dataset was divided into the training set and the test set according to the ratio of 7:3 to train and test the EPScGAN model. Experimental results show that, the mean Intersection over Union (mIoU) of the EPScGAN model on the test set is 78.12%, and compared with other 6 popular deep learning segmentation models, EPScGAN model has better segmentation performance.

Semi-supervised learning method for automatic nuclei segmentation using generative adversarial network

CHENG Kai, WANG Yan, LIU Jianfei

2020, 40(10):  2917-2922.  DOI: 10.11772/j.issn.1001-9081.2020020136

Asbtract ( )   PDF (3833KB) ( )  

References | Related Articles | Metrics

In order to reduce the dependence on the number of labeled images, a novel semi-supervised learning method was proposed for automatic segmentation of nuclei. Firstly, a novel Convolutional Neural Network (CNN) was used to extract the cell region from the background. Then, a confidence map for the input image was generated by the discriminator network via applying a full convolutional network. At the same time, the adversarial loss and the standard cross-entropy loss were coupled to improve the performance of the segmentation network. Finally, the labeled images and unlabeled images were combined with the confidence maps to train the segmentation network, so that the segmentation network was able to identify the nuclei in the extracted cell regions. Experimental results on 84 images (1/8 of the total images in the training set were labeled, and the rest were unlabeled) showed that the SEGmentation accuracy measurement (SEG) score of the proposed nuclei segmentation method achieved 77.9% and F1 score of the method was 76.0%, which were better than those of the method when using 670 images (all images in the training set were labeled).

Data science and technology

Scheduling method for big data tasks

LI Ziying, SHI Zhenguo

2020, 40(10):  2923-2928.  DOI: 10.11772/j.issn.1001-9081.2020030348

Asbtract ( )   PDF (903KB) ( )  

References | Related Articles | Metrics

Because the division and resource allocation of big data tasks lacks rationality in big data processing procedure, a scheduling method for big data tasks was proposed. First, in order to establish a reasonable management system of big data tasks and standardize the big data task processing flow, the scheduling theory was introduced to handle big data tasks. Then, based on the natures of big data tasks, the datasets were analyzed and handled, the decision table was introduced to perform attribute reduction, so as to reduce the data amount of big data analysis tasks and improve the big data analysis efficiency. Finally, the fuzzy comprehensive evaluation method was adopted, and the result of fuzzy comprehensive evaluation was used as the basis for task scheduling, thereby improving the rationality of task resource allocation. Experimental results on University of California Irvine (UCI) datasets show that the average prediction accuracy of the proposed scheduling algorithm is 7.42 percentage points higher than that of the Naive Bayes (NB) algorithm, 5.16 percentage points higher than that of the error Back Propagation (BP) algorithm, and 3.74 percentage points higher than that of the Root Mean Square Prop (RMSProp) algorithm. For datasets with a large number of features, the prediction accuracy of the proposed algorithm is significantly improved compared to those of other algorithms. Compared with Heterogeneous Critcal Path First Synthesis (HCPFS) algorithm and Heterogeneous Improved Priority List for Task Scheduling (HIPLTS) algorithm, the proposed algorithm has the average Scheduling Length Ratio (SLR) decreased by 12.14% and 4.56% respectively, and the average speedup ratio increased by 7.14% and 42.56% respectively, showing that the algorithm can effectively improve the efficiency of task scheduling in big data systems. Comprehensive analysis shows that the proposed algorithm performs well in prediction accuraing, and is efficient and reliable.

Network embedding based tenuous subgraph finding

SUN Heli, HE Liang, HE Fang, SUN Miaomiao, JIA Xiaolin

2020, 40(10):  2929-2935.  DOI: 10.11772/j.issn.1001-9081.2020020207

Asbtract ( )   PDF (1167KB) ( )  

References | Related Articles | Metrics

Concerning the high time and space complexity caused by using high-dimensional tenuous vectors to represent network information in tenuous subgraph finding problem, a Tenuous subGraph Finding (TGF) algorithm based on network embedding was proposed. Firstly, the network structure was mapped to the low-dimensional space by the network embedding method in order to obtain the low-dimensional vector representation of nodes. Then, the tenuous subset finding problem in the vector space was defined, and the tenuous subgraph finding problem was transformed into the tenuous subset finding problem. Finally, the sample points with lowest local density were searched iteratively and expanded to figure out the largest tenuous subset satisfying the given conditions. Experimental results on Synthetic_1000 dataset show that, the search efficiency of TGF algorithm is 1 353 times that of Triangle and Edge Reduction Algorithm (TERA) and 4 times of that Weight of K-hop (WK) algorithm, and it achieves better results in k-line, k-triangle and k-density indexes

Urban reachable region search based on time segment tree

SUN Heli, ZHANG Youyou, YANG Zhou, HE Liang, JIA Xiaolin

2020, 40(10):  2936-2941.  DOI: 10.11772/j.issn.1001-9081.2020020231

Asbtract ( )   PDF (1286KB) ( )  

References | Related Articles | Metrics

Aiming at the problem of reachable region search problem in urban computing, a method based on time segment tree was developed. In the method, a time segment tree structure was designed to store the local reachable regions, and a dynamic adaptive search algorithm was proposed, so as to improve the efficiency and accuracy of reachable region search. The method includes four steps. Firstly, the probability time weights of road segments were constructed on the basis of road speed distribution model and the trajectory data. Then, the short-term reachable regions were queried and stored by using the hierarchical skip list algorithm. After that, an efficient index structure for the hierarchical reachable region was built by the use of the time segment tree. Finally, the iterative search in the road network was carried out by using the time segment tree index, and the reachable region set was obtained. Extensive experiments were conducted on Beijing road network and taxi trajectory datasets. The results show that the proposed method improves the efficiency and accuracy by 18.6% and 25% respectively compared with the state-of-the-art Single-location reachability Query Maximum/minimum Bounding region search (SQMB) method.

Erasure code with low recovery-overhead in distributed storage systems

ZHANG Hang, LIU Shanzheng, TANG Dan, CAI Hongliang

2020, 40(10):  2942-2950.  DOI: 10.11772/j.issn.1001-9081.2020010127

Asbtract ( )   PDF (1250KB) ( )  

References | Related Articles | Metrics

Erasure code technology is a typical data fault tolerance method in distributed storage systems. Compared with multi-copy technology, it can provide high data reliability with low storage overhead. However, the high repair cost limits the practical applications of erasure code technology. Aiming at problems of high repair cost, complex encoding and poor flexibility of existing erasure codes, a simple-encoding erasure code with low repair cost - Rotation Group Repairable Code (RGRC) was proposed. According to RGRC, multiple strips were combined into a strip set at first. After that, the association relationship between the strips was used to hierarchically rotate and encode the data blocks in the strip set to obtain the corresponding redundant blocks. RGRC greatly reduced the amount of data needed to be read and transmitted in the process of single-node repair, thus saving a lot of network bandwidth resources. Meanwhile, RGRC still retained high fault tolerance when solving the problem of high repair cost of a single node. And, in order to meet the different needs of distributed storage systems, RGRC was able to flexibly weigh the storage overhead and repair cost of the system. Comparison experiments were conducted on a distributed storage system, the experimental analysis shows that compared with RS (Reed-Solomon) codes, LRC (Locally Repairable Codes), basic-Pyramid, DLRC (Dynamic Local Reconstruction Codes), pLRC (proactive Locally Repairable Codes), GRC (Group Repairable Codes) and UFP-LRC (Unequal Failure Protection based Local Reconstruction Codes), RGRC can reduce the repair cost of single node repair by 14%-61% through adding a small amount of storage overhead, and reduces the repair time by 14%-58%.

Course recommendation system based on R2 index and multi-objective differential evolution

HAO Qinxia

2020, 40(10):  2951-2959.  DOI: 10.11772/j.issn.1001-9081.2020010086

Asbtract ( )   PDF (1063KB) ( )  

References | Related Articles | Metrics

Aiming at the problem of the lack of accurate recommended and selected courses in the new form of higher education, a high-dimensional multi-objective evolutionary algorithm based course guidance and recommendation method was proposed. First, a multi-dimensional fact data warehouse model was designed to save storage space, and the related attributes in the data warehouse such as courses, students, teachers, course difficulty and course recommendation index were formally defined and stipulated. Second, a recommendation model based on high-dimensional R2-MODE (R2 based Multi-Objective Differential Evolution) algorithm was constructed, which improved the search ability in the high-dimensional complex space. Finally, the optimizations of 4 performances, the professionalism of the course teacher, the professional relevance of the course, the degree of the course difficulty and the comprehensive evaluation of the course, were achieved at the same time. Experimental results showed that the proposed algorithm improved the convergence by 50% compared with the reference point-based NSGA-Ⅲ (Third version of Non-dominated Sorting Genetic Algorithm), and had the increase of 5% in the distribution compared with the dominant relationship-based ε-MOEA (ε-dominance based Multi Objective Evolutionary Algorithm). The designed method had the best overall effect on the convergence and distribution of datasets. In the experiment, the accurate recommendation of courses according to the individual characteristics and wishes of students was successfully performed by using the proposed algorithm. The proposed algorithm provided the necessary theoretical support for the accurate guidance and recommendation of course selection on the network platform, and a new method for intelligent course selection.

Cyber security

Generative adversarial network-based system log-level anomaly detection algorithm

XIA Bin, BAI Yuxuan, YIN Junjie

2020, 40(10):  2960-2966.  DOI: 10.11772/j.issn.1001-9081.2020020270

Asbtract ( )   PDF (1412KB) ( )  

References | Related Articles | Metrics

To solve the problems of small number of anomaly samples and inefficient feedback of anomalies in the anomaly detection tasks of large-scale software system, a log-level anomaly detection algorithm based on Generative Adversarial Network (GAN) and attention mechanism. First, the unstructured logs were converted into structured events through the log templates, and each event included timestamps, signature and parameters. Second, through sliding window method, the sequence of the parsed events were divided into patterns, and the real training dataset was comprised combination of the divided event patterns and the corresponding following events. Third, the real event patterns were used as the training samples to train the attention mechanism-based GAN, and the Recurrent Neural Network (RNN) based generator was trained through the adversarial learning mechanism until it converged. Finally, through the input flow event pattern, the generator generated the possibility distribution of normal and abnormal events based on the previous pattern. When the threshold was set, whether the specific log of next moment is a normal event or an abnormal event was determined automatically. Experimental results show that the proposed anomaly detection algorithm, which uses a gated recurrent unit network as the attention weight and a Long Short-Term Memory (LSTM) network to fit event patterns, has a 21.7% increase in precision compared to the algorithm only using the gated recurrent unit network. In addition, compared to the log-level anomaly detection algorithm LogGAN, the proposed algorithm improves the precision of anomaly detection by 7.8% over the performance of LogGAN.

Integral attack on PICO algorithm based on division property

LIU Zongfu, YUAN Zheng, ZHAO Chenxi, ZHU Liang

2020, 40(10):  2967-2972.  DOI: 10.11772/j.issn.1001-9081.2019122228

Asbtract ( )   PDF (810KB) ( )  

References | Related Articles | Metrics

PICO proposed in recent years is a bit-based ultra lightweight block cipher algorithm. The security of this algorithm to resist integral cryptanalysis was evaluated. Firstly, by analyzing the structure of PICO cipher algorithm, a Mixed-Integer Linear Programming (MILP) model of the algorithm was established based on division property. Then, according to the set constraints, the linear inequalities were generated to describe the propagation rules of division property, and the MILP problem was solved with the help of the mathematical software, the success of constructing the integral distinguisher was judged based on the objective function value. Finally, the automatic search of integral distinguisher of PICO algorithm was realized. Experimental results showed that, the 10-round integral distinguisher of PICO algorithm was searched, which is the longest one so far. However, the small number of plaintexts available is not conducive to key recovery. In order to obtain better attack performance, the searched 9-round distinguisher was used to perform 11-round key recovery attack on PICO algorithm. It is shown that the proposed attack can recover 128-bit round key, the data complexity of the attack is 2^63.46, the time complexity is 2⁷⁶ 11-round encryptions, and the storage complexity is 2²⁰.

Slow HTTP DoS attack detection method based on one-dimensional convolutional neural network

CHEN Yi, ZHANG Meijing, XU Fajian

2020, 40(10):  2973-2979.  DOI: 10.11772/j.issn.1001-9081.2020020172

Asbtract ( )   PDF (1225KB) ( )  

References | Related Articles | Metrics

In order to solve the problem that the accuracy of Slow HTTP Denial of Service (SHDoS) attack traffic detection decreases when the attack frequency changes, a method of SHDoS attack traffic detection method based on one-dimensional Convolutional Neural Network (CNN) was proposed. First, the message sampling and data stream extraction were performed on three types of SHDoS attack traffic under multiple attack frequencies by the method. Then, a data stream conversion algorithm was designed to convert the collected attack data streams into one-dimensional sequences and remove the duplicated sequences. Finally, a one-dimensional CNN was used to construct a classification model. The model was used to extract sequence fragments through the convolution kernels, and the local patterns of attack samples were learned from fragments. Therefore, the model would have the ability to detect attack traffic with multiple attack frequencies. Experimental results show that, compared with the classification models based on Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM) network, and Bidirectional LSTM (Bi-LSTM) network respectively, the proposed model has advantages in detection performance on unknown frequency samples, and has the accuracy and precision reached 96.76% and 94.13% respectively on the validation set. It can be seen that the proposed method can meet the needs of detecting SHDoS traffic with different attack frequencies.

Federated security tree algorithm for user privacy protection

ZHANG Junru, ZHAO Xiaoyan, YUAN Peiyan

2020, 40(10):  2980-2985.  DOI: 10.11772/j.issn.1001-9081.2020030332

Asbtract ( )   PDF (1608KB) ( )  

References | Related Articles | Metrics

Aiming at the problems of low accuracy and low operation efficiency of federated learning algorithm in user behavior prediction, a loss-free Federated Learning Security tree (FLSectree) algorithm was proposed. Firstly, through the derivation of the loss function, its first partial derivative and second partial derivative were proved to be sensitive data, and the optimal split point after encryption was returned by scanning and splitting the feature index sequence, so as to protect the sensitive data from being disclosed. Then, by updating the instance space, the splitting was continued and the next best split point was found until the termination condition was satisfied. Finally, the results of training were used to obtain local algorithm parameters for each participant. Experimental results show that the FLSectree algorithm can effectively improve the accuracy and the training efficiency of user behavior prediction algorithm under the premise of protecting the data privacy. Compared with the SecureBoost algorithm in Federated AI Technology Enabler (FATE) framework of federated learning, FLSectree algorithm has the user behavior prediction accuracy increased by 9.09% and has the operation time reduced by 87.42%, and the training results are consistent with centralized Xgboost algorithm.

Order preserving encryption scheme of nonlinear mapping based on random function

XU Yansheng, ZHANG Youjie

2020, 40(10):  2986-2991.  DOI: 10.11772/j.issn.1001-9081.2020020167

Asbtract ( )   PDF (1061KB) ( )  

References | Related Articles | Metrics

To solve the problem that the existing order preserving encryption schemes are difficult to give consideration to security, efficiency and ease of use at the same time, an order preserving encryption scheme of non-linear mapping based on random function was proposed. In the scheme, the plaintext space was considered as an increasing arithmetic sequence, and each element of the sequence was mapped to a separate ciphertext space based on the key. The key was generated by a random number generating function with non-uniform distribution, and the ciphertext space was constructed by a computer program. During encrypting, the value randomly selected from the corresponding ciphertext space was able to be used as the ciphertext. Analysis and experimental results show that the proposed scheme achieves INDistinguishability under Ordered Chosen Plaintext Attack (IND-OCPA) safety and can effectively prevent statistical attacks; it has the average encryption time per 100 000 data of from 30 ms to 50 ms, resulting in high encryption efficiency; the complex parameter presets are not required in the scheme, and the scheme can be implemented in any computer language, so that it is easy to use.

Blockchain based efficient anonymous authentication scheme for IOV

CHEN Weiwei, CAO Li, SHAO Changhong

2020, 40(10):  2992-2999.  DOI: 10.11772/j.issn.1001-9081.2020020211

Asbtract ( )   PDF (1304KB) ( )  

References | Related Articles | Metrics

In order to solve the problems of low efficiency of centralized authentication and poor privacy protection in Internet of Vehicles (IOV), an efficient anonymous authentication scheme based on blockchain technology was proposed. According to the IOV's characteristics of openness, self-organization and fast movement, the tamper-proof and distributed features of blockchain technology were used to realize the generation and blockchain storing of temporary identities of the vehicles. Smart contract was implemented to make efficient anonymous two-way identity authentication while vehicles communicated with each other. Experimental results show that, in terms of authentication efficiency, the proposed scheme has the anonymous authentication with slower time delay growth and higher efficiency compared with traditional Public Key Infrastructure (PKI) authentication and identity authentication scheme with pseudonym authorization; in terms of safety performance, the temporary identity stored in the blockchain has characteristics of non-tampering, nondenying and traceability. In this scheme, the malicious vehicle identity and authority can be traced back and controlled respectively, and the public-key cryptography and digital signature technology ensure the confidentiality and integrity of communication data.

Network and communications

Unambiguous tracking method of AltBOC (15, 10) signal based on combined correlation function

YUAN Zhixin, ZHOU Yanling

2020, 40(10):  3000-3005.  DOI: 10.11772/j.issn.1001-9081.2020030289

Asbtract ( )   PDF (1183KB) ( )  

References | Related Articles | Metrics

Aiming at the tracking ambiguity of Alternate Binary Offset Carrier (AltBOC) (15, 10) signal caused by the multi-peak of autocorrelation function, an unambiguous tracking method based on combined correlation function was proposed. Firstly, two special local reference signals were designed in this method, then the correlation operations between the local reference signals and the received signals were realized through the code tracking loop to obtain two cross-correlation functions, and finally the ambiguity was eliminated through the unambiguous correlation function which is the multiplication of the two obtained cross-correlation functions. The tracking accuracy under thermal noise and anti-multipath interference performance of the proposed method were studied. The results show that:1) the proposed method realizes the unambiguous tracking; 2) the tracking accuracy of the proposed method is between Binary Phase Shift Keying Like (BPSK Like) method and Pseudo Correlation Function (PCF) method, and the tracking accuracy of the proposed method is nearly close to that of PCF method when the noise ratio is greater than 40 dB·Hz; 3) when the multipath delay is greater than 0.1 code chip, the proposed method has the multipath error significantly smaller than that of BPSK Like method and PCF method.

Ray tracking acceleration method based on combination of indoor dynamic and static divisions

HUANG Yihang, JIANG Hong, HAN Bin

2020, 40(10):  3006-3012.  DOI: 10.11772/j.issn.1001-9081.2020020200

Asbtract ( )   PDF (6394KB) ( )  

References | Related Articles | Metrics

Channel modeling of the closed environment plays an important role in many application scenarios. When there are many obstacles in the space, the traditional ray tracing algorithm has the problem of too many times of finding intersection points in the calculation process, which makes the algorithm calculation efficiency low. Therefore, a ray tracing acceleration method based on space division was proposed. In the method, according to the distribution of objects in three-dimensional space, the static and dynamic space division acceleration methods were combined reasonably, so as to greatly reduce the number of finding intersection points between rays and objects in space, and improve the calculation efficiency of the algorithm. Simulation analysis shows that in the three-dimensional environment with the same prediction accuracy, compared with the original algorithm, the ray tracing algorithm using static space division has the calculation efficiency improved by at least 50.2% as the division level is improved; and compared with the algorithm which only uses static space division, the acceleration method based on the combination of static and dynamic space divisions has the calculation efficiency improved by at least 8.9% on the basis of the improvement above.

Computer software technology

Markov process-based availability modeling and analysis method of IaaS system

YANG Shenshen, WU Huizhen, ZHUANG Lili, LYU Hongwu

2020, 40(10):  3013-3018.  DOI: 10.11772/j.issn.1001-9081.2019122245

Asbtract ( )   PDF (1088KB) ( )  

References | Related Articles | Metrics

Concerning the problem that existing availability models of Infrastructure as a Service (IaaS) are difficult to calculate the probability of the existence of multiple available Physical Machines (PMs), a new availability analysis method based on Markov process was proposed for IaaS clouds. Firstly, the computing resources were divided into three types:hot PM, warm PM and cold PM. Then, the impact of availability was modeled by combining the corresponding stages of the resource allocation process, separately generating three kinds of allocation sub-models. These sub-models cooperated with each other through the transformation relationships of different types of computing resources, so as to construct the overall model of the system. After that, the availability model was solved by equations constructed based on Markov process. Finally, the proposed analysis model was verified with a practical example, and the key factors such as PM transition rate were analyzed. Experimental results show that, increasing the number of PMs, especially cold PMs helps to improve the availability of IaaS. The proposed method can be used to estimate the probability of the existence of one or multiple available PMs.

Truthful mechanism for crowdsourcing task assignment in social network

QIN Haiyan, ZHANG Yonglong, LI Bin

2020, 40(10):  3019-3024.  DOI: 10.11772/j.issn.1001-9081.2020020174

Asbtract ( )   PDF (922KB) ( )  

References | Related Articles | Metrics

There are more and more macrotasks on the crowdsourcing platforms. Macrotasks require the professional skills of the workers and the collective contributions of the teams. Social networks provide a platform for cooperation among social workers. In fact, there are few studies paying close attention to the social network between crowdsourcing workers. The problem of task assignment in crowdsourcing is NP-hard and the participants may misreport their prices in order to gain more utilities. Therefore, a Truthful Mechanism for Crowdsourcing task assignment in Social Network (TMC-SN) was proposed. The problem of crowdsourcing task assignment in social network was modeled as an auction, where the task requester is the buyer, the workers are the sellers, and the crowdsourcing platform is served as an auctioneer. In order to find the most suitable team, TMC-SN measured worker fitness to the team from both marginal contribution and team cohesion. Theoretical analysis verifies that TMC-SN has economic properties such as truthfulness, individual rationality, and budget balance. Experimental results show that TMC-SN has certain advantages in social welfare, and can improve the utilities of workers.

Consistency analysis method of software design and implementation based on control flow

ZHANG Jiaqi, MU Yongmin, ZHANG Zhihua

2020, 40(10):  3025-3033.  DOI: 10.11772/j.issn.1001-9081.2020030311

Asbtract ( )   PDF (1635KB) ( )  

References | Related Articles | Metrics

The current consistency detection methods of software design and implementation require a large number of template sets and are difficult to generalize. In order to solve these problems, a consistency analysis method of software design and implementation based on control flow was proposed. Firstly, the pseudocode of the design document and the source code of the program were converted into the intermediate representations with the same features, and the design feature and the implementation feature were respectively extracted from the intermediate representations. The features include the function call relationship which can reflect the system structure and the control flow information which can reflect the internal structure of the function. Then, the design feature model and the implementation feature model were respectively established according to the design feature and the implementation feature. Finally, the similarity of the feature model was measured by calculating the feature similarity, so as to obtain the consistency detection result. Experimental results show that this method can correctly detect the inconsistent function call relationship when the function call relationship realized by the software is inconsistent with the design, and can correctly detect the inconsistency of the internal structure of the function when the function call relationship realized by the software is consistent with the design, with the accuracy reached 92.85%. This method can effectively obtain the consistency detection results without any template set, and has superior generality.

Virtual reality and multimedia computing

Speaker recognition in strong noise environment based on auditory cortical neuronal receptive field

NIU Xiaoke, HUANG Yixin, XU Huaxing, JIANG Zhenyang

2020, 40(10):  3034-3040.  DOI: 10.11772/j.issn.1001-9081.2020020272

Asbtract ( )   PDF (1737KB) ( )  

References | Related Articles | Metrics

Aiming at the problem that speaker recognition is susceptible to environmental noise, a new voiceprint extraction method was proposed based on the spatial-temporal filtering mechanism of Spectra-Temporal Receptive Field (STRF) of biological auditory cortex neurons. In the method, the quadratic characteristics were extracted from the auditory scale-rate map based on STRF, and the traditional Mel-Frequency Cepstral Coefficient (MFCC) was combined to obtain the voiceprint features with strong tolerance to environmental noise. Using Support Vector Machine (SVM) as feature classifier, the testing results on speech data with different Signal-to-Noise Ratios (SNR) showed that the STRF-based features were more robust to noise than MFCC coefficient, but had lower recognition accuracy; the combined features improved the accuracy of speech recognition and had good robustness to noise. The results verify the effectiveness of the proposed method in speaker recognition under strong noise environment.

Image super-resolution reconstruction method combining perceptual edge constraint and multi-scale fusion network

OUYANG Ning, WEI Yu, LIN Leping

2020, 40(10):  3041-3047.  DOI: 10.11772/j.issn.1001-9081.2020020185

Asbtract ( )   PDF (4821KB) ( )  

References | Related Articles | Metrics

Aiming at the problems that the image super-resolution reconstruction model requires a large number of parameters to capture the statistical relationship between Low-Resolution (LR) images and High-Resolution (HR) images, and the use of network models optimized by L₁ or L₂ loss cannot effectively recover the high-frequency details of the images, an image super-resolution reconstruction method combining perceptual edge constraint and multi-scale fusion network was proposed. Based on the idea from coarse to fine, a two-stage network model was designed in this method. At the first stage, Convolutional Neural Network (CNN) was used to extract image features and upsample the image features to the HR size in order to obtain rough features. At second stage, multi-scale estimation was used to gradually approximate the low-dimensional statistical model to the high-dimensional statistical model. The rough features output at the first stage were used as the input to extract the multi-scale features of the image, and the features of different scales were gradually fused together through the attention fusion module in order to refine the features extracted at the first stage. At the same time, a class of richer convolutional features was introduced for edge detection and used as the perceptual edge constraint to optimize the network, so as to better recover the high-frequency details of the images. Experimental results on benchmark datasets such as Set5, Set14 and BSDS100 show that compared with the existing CNN-based super-resolution reconstruction methods, the proposed method not only reconstructs sharper edges and textures, but also achieves certain improvements in Peak Signal-to-Noise Ratio (PSNR) and Structural SIMilarity index (SSIM) when magnification factor is 3 and 4.

Mixed-order channel attention network for single image super-resolution reconstruction

YAO Lu, SONG Huihui, ZHANG Kaihua

2020, 40(10):  3048-3053.  DOI: 10.11772/j.issn.1001-9081.2020020281

Asbtract ( )   PDF (3787KB) ( )  

References | Related Articles | Metrics

For the current channel attention mechanism used for super-resolution reconstruction, there are problems that the attention prediction destroys the direct corresponding relationship between each channel and its weight and the mechanism only considers the first-order or second-order channel attention without comprehensive consideration of the advantage complementation. Therefore, a mixed-order channel attention network for image super-resolution reconstruction was proposed. First of all, by using the local cross-channel interaction strategy, increase and reduction in channel dimension used by the first-order and second-order channel attention models were changed into a fast one-dimensional convolution with kernel k, which not only makes the channel attention prediction more direct and accurate but makes the resulting model simpler than before. Besides, the improved first and second-order channel attention models above were adopted to comprehensively take the advantages of channel attentions of different orders, thus improving network discrimination. Experimental results on the benchmark datasets show that compared with the existing super-resolution algorithms, the proposed method has the best recovered texture details and high frequency information of the reconstructed images and the Perceptual Indictor (PI) on Set5 and BSD100 datasets are increased by 0.3 and 0.1 on average respectively. It shows that this network is more accurate in predicting channel attention and comprehensively uses channel attentions of different orders, so as to improve the performance.

Compressed sensing magnetic resonance imaging based on deep priors and non-local similarity

ZONG Chunmei, ZHANG Yueqin, CAO Jianfang, ZHAO Qingshan

2020, 40(10):  3054-3059.  DOI: 10.11772/j.issn.1001-9081.2020030285

Asbtract ( )   PDF (1058KB) ( )  

References | Related Articles | Metrics

Aiming at the problem of low reconstruction quality of the existing Compressed Sensing Magnetic Resonance Imaging (CSMRI) algorithms at low sampling rates, an imaging method combining deep priors and non-local similarity was proposed. Firstly, a deep denoiser and Block Matching and 3D filtering (BM3D) denoiser were used to construct a sparse representation model that can fuse multiple priori knowledge of images. Secondly, the undersampled k-space data was used to construct a compressed sensing magnetic resonance imaging optimization model. Finally, an alternative optimization method was used to solve the constructed optimization problem. The proposed algorithm can not only use the deep priors through the deep denoiser, but also use the non-local similarity of the image through the BM3D denoiser to reconstruct the image. Compared with the reconstruction algorithms based on BM3D, experimental results show that the proposed algorithm has the average peak signal-to-noise ratio of reconstruction increased about 1 dB at the sampling rates of 0.02, 0.06, 0.09 and 0.13. Compared with the existing MRI algorithm WaTMRI (Magnetic Resonance Imaging with Wavelet Tree sparsity),DLMRI (Dictionary Learning for Magnetic Resonance Imaging), DUMRI-BM3D (Magnetic Resonance Imaging based on Dictionary Updating and Block Matching and 3D filtering), etc, the images reconstructed by the proposed algorithm contain a lot of texture information, which are the closest to the original images.

Undersampled brain magnetic resonance image reconstruction method based on convolutional neural network

DU Nianmao, XU Jiachen, XIAO Zhiyong

2020, 40(10):  3060-3065.  DOI: 10.11772/j.issn.1001-9081.2020030344

Asbtract ( )   PDF (3713KB) ( )  

References | Related Articles | Metrics

Aiming at the problem that current deep learning based undersampled Magnetic Resonance (MR) image reconstruction methods mainly focus on the single slice reconstruction and ignore the data redundancy between adjacent slices, a Hybrid Cascaded Convolutional Neural Network (HC-CNN) was proposed for undersampled multi-slice brain MR image reconstruction. First, the traditional reconstruction method was extended to a deep learning based reconstruction model, and the traditional iterative reconstruction framework was replaced by a cascaded convolutional neural network. Then, in each iterative reconstruction, a 3D convolution module and a 2D convolution module were used to learn the data redundancy between adjacent slices and inside a single slice, respectively. Finally, Data Consistency (DC) module was used in each iteration to maintain the data fidelity of the reconstructed image in k-space. The simulation results on a single-coil brain MR image dataset show that compared with the reconstruction methods based on single slice reconstruction, the proposed method has the Peak Signal-to-Noise Ratio (PSNR) value at 4×acceleration factor increased by 1.75 dB averagely and the PSNR value at 6×acceleration factor increased by 2.57 dB averagely. At the same time, the image reconstruction time for a single slice by the proposed method is 15.4 ms. Experimental results show that the proposed method can not only effectively utilize the data redundancy between slices and reconstruct higher-quality images, but also has a higher real-time performance.

Rail surface defect detection method based on background differential with defect proportion limitation

CAO Yiqin, LIU Longbiao

2020, 40(10):  3066-3074.  DOI: 10.11772/j.issn.1001-9081.2020030337

Asbtract ( )   PDF (3568KB) ( )  

References | Related Articles | Metrics

Aiming at the characteristics of rail surface images such as uneven illumination, limited discernible features, low contrast and changeable reflection characteristics, a background differential rail surface defect detection method based on defect proportion limitation was proposed. The method mainly includes five steps:pre-processing of rail surface images, background modeling and difference, defect proportion limitation filtering, maximum entropy threshold segmentation of defect proportion limitation and connected area labeling. Firstly, the column grayscale mean and median of the rail surface image were combined to perform the rapid background modeling, and the difference operation was carried out to the pre-processed image and the background image. Secondly, the feature with low defect proportion in the rail surface image was used to truncate the upper threshold limit of the defect proportion in order to enhance the contrast of the difference image. Thirdly, the maximum entropy threshold segmentation was improved by using this feature, the global variable weighting of the target entropy was carried out by using the adaptive weighting factor, and an appropriate threshold was selected to maximize the entropy value, so as to reduce the interference of noises such as shadow and rust while retaining the real defects. Finally, the connected area labeling method was used to perform the statistics of the defect areas in the segmented binary image, and the area with defect area lower than the rail damage standard was determined as the noise and removed, so as to realize the rail surface defect detection. Simulation results show that the new method can detect rail surface defects well, and its results have the recall rate, precision rate and weighted harmonic mean of 94.19%, 88.34% and 92.96% respectively, and the average mis-classification error of 0.006 4, so that the method has certain practical value.

Multiple aerial infrared target tracking method based on multi-feature fusion and hierarchical data association

YANG Bo, LIN Suzhen, LU Xiaofei, LI Dawei, QIN Pinle, ZUO Jianhong

2020, 40(10):  3075-3080.  DOI: 10.11772/j.issn.1001-9081.2020030320

Asbtract ( )   PDF (1977KB) ( )  

References | Related Articles | Metrics

An online multiple target tracking method for the aerial infrared targets was proposed based on the hierarchical data association to solve the tracking difficulty caused by the high similarity, large number and large false detections of the targets in star background. Firstly, according to the characteristics of the infrared scene, the location features, gray features and scale features of the targets were extracted. Secondly, the above three features were combined to calculate the preliminary relationship between the targets and the trajectories in order to obtain the real targets. Thirdly, the obtained real targets were classified according to their scales. The large-scale target data association was calculated by adding three features of appearance, motion and scale. The small-scale target data association was calculated by multiplying the two features of appearance and motion. Finally, the target assignment and trajectory updating were performed to the two types of targets respectively according to the Hungarian algorithm. Experimental results in a variety of complex conditions show that:compared with the online tracking method only using motion features, the proposed method has the tracking accuracy improved by 12.6%; compared with the method using multi-feature fusion, the hierarchical data correlation of the proposed method not only improves the tracking speed, but also increases the tracking accuracy by 19.6%. In summary, this method not only has high tracking accuracy, but also has good real-time performance and anti-interference ability.

Frontier & interdisciplinary applications

Heterogeneous sensing multi-core scheduling method based on machine learning

AN Xin, KANG An, XIA Jinwei, LI Jianhua, CHEN Tian, REN Fuji

2020, 40(10):  3081-3087.  DOI: 10.11772/j.issn.1001-9081.2020010118

Asbtract ( )   PDF (1048KB) ( )  

References | Related Articles | Metrics

Heterogeneous multi-core processor is the mainstream solution for modern embedded systems now. Good online mapping or scheduling approaches play important roles in improving their advantages of high performance and low power consumption. To deal with the problem of dynamic mapping and scheduling of applications on heterogeneous multi-core processing systems, a dynamic mapping and scheduling solution was proposed to effectively determine remapping time in order to maximize the system performance by using the machine learning based detection technology of quickly and accurately evaluating program performance and program behavior phase change. In this solution, by carefully selecting the static and dynamic features of processing cores and programs to running to effectively detect the difference in computing power and workload running behaviors brought by heterogeneous processing, a more accurate prediction model was built. At the same time, by introducing phase detection technology, the number of online mapping computations was reduced as much as possible, so as to provide more efficient scheduling scheme. Finally, the effectiveness of the proposed scheduling scheme was verified on the SPLASH-2 dataset. Experimental results showed that, compared to the Completely Fair Scheduler (CFS) of Linux, the proposed method achieved about 52% computing performance gains and 9.4% improvement on CPU resource utilization rate. It shows that the proposed method has excellent performance in system computing performance and processor resource utilization, and can effectively improve the dynamic mapping and scheduling effect of applications of heterogeneous multi-core systems.

Early diagnosis and prediction of Parkinson's disease based on clustering medical text data

ZHANG Xiaobo, YANG Yan, LI Tianrui, LU Fan, PENG Lilan

2020, 40(10):  3088-3094.  DOI: 10.11772/j.issn.1001-9081.2020030359

Asbtract ( )   PDF (1270KB) ( )  

References | Related Articles | Metrics

In view of the problem of the early intelligent diagnosis for Parkinson's Disease (PD) which occurs more common in the elderly, the clustering technologies based on medical detection text information data were proposed for the analysis and prediction of PD. Firstly, the original dataset was pre-processed to obtain effective feature information, and these features were respectively reduced to eight dimensional spaces with different dimensions by Principal Component Analysis (PCA) method. Then, five traditional classical clustering models and three different clustering ensemble methods were respectively used to cluster the data of eight dimensional spaces. Finally, four clustering performance indexes were selected to predict PD subject with dopamine deficiency as well as healthy control and Scans Without Evidence of Dopamine Deficiency (SWEDD) PD subject. The simulation results show that the clustering accuracy of Gaussian Mixture Model (GMM) reaches 89.12% when the value of PCA feature dimension is 30, the clustering accuracy of Spectral Clustering (SC) is 61.41% when the PCA feature dimension value is 70, and the clustering accuracy of Meta-CLustering Algorithm (MCLA) achieves 59.62% when the PCA feature dimension value is 80. The comparative experiments results show that GMM has the best clustering effect in the five classical clustering methods when the PCA feature dimension value is less than 40 and MCLA has the excellent clustering performance among the three clustering ensemble methods for different feature dimensions, which thereby provides the technical and theoretical supports for the early intelligent auxiliary diagnosis of PD.

Method of semantic entity construction and trajectory control for UAV electric power inspection

REN Na, ZHANG Nan, CUI Yan, ZHANG Rongxue, PANG Xinfu

2020, 40(10):  3095-3100.  DOI: 10.11772/j.issn.1001-9081.2020020198

Asbtract ( )   PDF (2216KB) ( )  

References | Related Articles | Metrics

The reasonable control of trajectory is an important factor affecting intelligent decision-making of Unmanned Aerial Vehicle (UAV). Focusing on the local observability and the complexity of upper air of mission environment, a method of semantic entity construction and trajectory control for UAV electric power inspection was proposed. Firstly, a spatial topology network based on entity knowledge of electric power inspection field was built, and the semantic trajectory sequence network about position nodes and its semantic interfaces were generated. Then, based on the result set of similarity measure of spatial topology structures, the security licensing mechanism and reinforcement learning based trajectory control strategy were proposed to realize the UVA electric power inspection on the basis of consensus concept connotation and position structure. Experimental results show that for an example of UAV electric power inspection, the optimal strategy obtained by the proposed method can satisfy the maximum robust performance, and at the same time, the fitness of the target network can stably converge and the physical area coverage is higher than 95% through the reinforcement learning of this method, so that the method provides flight basis for the decision-making of UVA electric power inspection tasks.