The print defect detection methods based on traditional image processing technology have poor robustness and the object detection methods based on deep learning are not completely suitable for the detection tasks of print defects. In order to solve the problems above， the comparison ideas in template matching method were combined with the semantic features in deep learning， and a Deep Comparison Network （CoNet） used for the detection tasks of print defects was proposed. Firstly， the Deep Comparison Module （DCM） adopting Siamese structure was proposed to mine the semantic relationship between the detection image and the reference image through extracting and fusing the feature maps of them in the semantic space. Then， based on the feature pyramid structure with asymmetric dual channels， the Multi-scale Change Detection Module （MsCDM） was proposed to locate and classify print defects. On the public printed circuit board defect dataset DeepPCB and dataset of Lijin defects， the average values of mean Average Precision （mAP） of CoNet are 99.1% and 69.8% respectively， compared with the two baseline models Max-Pooling Group Pyramid Pooling （MP-GPP） and Change-Detection Single Shot Detector （CD-SSD）， which are increased by 0.4， 3.5 percentage points and 0.7， 2.4 percentage points respectively， and the detection accuracy of CoNet is higher. Besides， when the resolution of input image is 640×640， the average time consumption of CoNet is 35.7 ms， showing that it can absolutely meet the real-time requirements of industrial detection tasks.

Aiming at the low detection precision of strip steel surface defects in actual scenarios， which is prone to missed detection and false detection， a YOLOv5-CFD model consisted of CSPDarknet53， Frequency channel attention Network （FcaNet） and Decoupled head was constructed to detect strip steel defects more accurately. Firstly， Fuzzy C-Means （FCM） algorithm was used to cluster anchor boxes in NEU-DET hot-rolling strip steel surface defect detection dataset published by Northeastern University to optimize the matching degree between the prior box and the ground-truth box. Secondly， in order to extract the rich detailed information of the target area， the frequency domain channel attention module FcaNet （Frequency channel attention Network） was added to the original YOLOv5 algorithm. Finally， the decoupled head was used to separate the classification and regression tasks. Experimental results on NEU-DET dataset show that with introducing a small number of parameters to the original YOLOv5 algorithm， the improved YOLOv5 algorithm has the detection precision increased by 4.2 percentage points， the detection mean Average Precision （mAP） of 85.5%； and the detection speed reaches 27.71 Frames Per Second （FPS）， which is not much different from the original YOLOv5 so that YOLOv5-CFD can meet the real-time detection requirements.

Aiming at the problem that current detectors based on deep learning cannot effectively detect objects with irregular shapes or large differences between length and width, based on the traditional Faster Region-based Convolutional Neural Network (Faster R-CNN) algorithm, an improved two-stage object detection framework named Accurate R-CNN was proposed. First of all, a novel Intersection over Union (IoU) metric-Effective Intersection over Union (EIoU) was proposed to reduce the proportion of redundant bounding boxes in the training data by using the centrality weight. Then, a context related Feature Reassignment Module (FRM) was proposed to re-encode the features by the remote dependency and local context information of objects, so as to make up for the loss of shape information in the pooling process. Experimental results show that on the Microsoft Common Objects in COntext (MS COCO) dataset, for the bounding box detection task, when using Residual Networks (ResNets) with two different depths of 50 and 101 as the backbone networks, Accurate R-CNN has the Average Precision (AP) improvements of 1.7 percentage points and 1.1 percentage points respectively compared to the baseline model Faster R-CNN, which are significantly than those of the detectors based on mask with the same backbone networks. After adding mask branch, for the instance segmentation task, when ResNets with two different depths are used as the backbone networks, the mask Average Precisions of Accurate R-CNN are increased by 1.2 percentage points and 1.1 percentage points respectively compared with Mask Region-based Convolutional Neural Network (Mask R-CNN). The research results illustrate that compared to the baseline model, Accurate R-CNN achieves better performance on different datasets and different tasks.

As the turbofan engine is one of the core equipment in the field of aerospace, its health condition determines whether the aircraft could work stably and reliably. And the prediction of the Remaining Useful Life (RUL) of turbofan engine is an important part of equipment monitoring and maintenance. In view of the characteristics such as complicated operating conditions, diverse monitoring data, and long time span existing in the turbofan engine monitoring process, a remaining useful life prediction model for turbofan engines integrating Genetic Algorithm-based Selective ENsembling (GASEN) and Temporal Convolutional Network (TCN) (GASEN-TCN) was proposed. Firstly, TCN was used to capture the inner relationship between data under long span, so as to predict the RUL. Then, GASEN was applied to ensemble multiple independent TCNs for enhancing the generalization performance of the model. Finally, the proposed model was compared with the popular machine learning methods and other deep neural networks on the general Commercial Modular Aero-Propulsion System Simulation (C-MAPSS) dataset. Experimental results show that, the proposed model has higher prediction accuracy and lower prediction error than the state-of-the-art Bidirectional Long-Short Term Memory (Bi-LSTM) network under many different operating modes and fault conditions. Taking FD001 dataset as an example:on this dataset, the Root Mean Square Error (RMSE) of the proposed model is 17.08% lower than that of Bi-LSTM, and the relative accuracy (Accuracy) of the proposed model is 12.16% higher than that of Bi-LSTM. It can be seen that the proposed model has considerable application prospect in intelligent overhaul and maintenance of equipment.

The traditional Average TimeSynch (ATS) for APron Sensor Network (APSN) has slow convergence and low algorithm efficiency due to its distributed iteration characteristics, based on the principle that the algebraic connectivity affects the convergence speed of the consensus algorithm, a Fast Convergence Average TimeSynch (FCATS) was proposed. Firstly, the virtual link was added between the two-hop neighbor nodes in APSN to increase the network connectivity. Then, the relative clock skew, logical clock skew and offset of the node were updated based on the information of the single-hop and two-hop neighbor nodes. Finally, according to the clock parameter update process, the consensus iteration was performed. The simulation results show that FCATS can be converged after the consensus iteration. Compared with ATS, it has the convergence speed increased by about 50%. And under different topological conditions, the convergence speed of it can be increased by more than 20%. It can be seen that the convergence speed is significantly improved.

The manual segmentation of chromosome images is time-consuming and laborious, and the accuracy of current automatic segmentation methods is not poor. Therefore, based on improved Mask R-CNN (Mask Region-based Convolutional Neural Network), a chromosome image segmentation framework named Mask Oriented R-CNN (Mask Oriented Region-based Convolutional Neural Network) was proposed, which introduced orientation information to perform instance segmentation of chromosome images. Firstly, the regression branch of oriented bounding boxes was added to predict the compact bounding boxes and obtain orientation information. Secondly, a novel Intersection-over-Union (IoU) metric called AwIoU (Angle-weighted Intersection-over-Union) was proposed to improve the criterion of redundant bounding boxes by combining the relationship between the orientation information and edges. Finally, the oriented convolutional path structure was realized to reduce the interference in mask prediction by copying the path of mask branch and selecting the training path according to the orientation information of the instances. Experimental results show that compared with the baseline model Mask R-CNN, Mask Oriented R-CNN has the mean average precision increased by 10.22 percentage points when the IoU threshold is 0.5, and the mean metric increased by 4.91 percentage points when the IoU threshold is from 0.5 to 0.95. Experimental results show that the Mask Oriented R-CNN framework achieves better segmentation results than the baseline model in chromosome image segmentation, which is helpful to achieve automatic segmentation of chromosome images.

The traditional handcrafted features rely heavily on the appearance characteristics of pedestrians and the deep convolution feature is a high-dimensional feature, so, it will consume a lot of time and memory when the feature is directly used to match the image. Moreover, features from higher levels are easily affected by human pose or background clutter. Aiming at these problems, a method based on deep multi-view feature distance learning was proposed. Firstly, a new feature to improve and integrate the convolution feature of the deep region was proposed. The convolution feature was processed by the sliding frame technique, and the integration feature of low-dimensional deep region with the dimension equal to the number of convolution layer channels was obtained. Secondly, from the perspectives of the deep regional integration feature and the handcrafted feature, a multi-view feature distance learning algorithm was proposed by utilizing the cross-view quadratic discriminant analysis method. Finally, the weighted fusion strategy was used to accomplish the collaboration between handcrafted features and deep convolution features. Experimental results show that the Rank1 value of the proposed method reaches 80.17% and 75.32% respectively on the Market-1501 and VIPeR datasets; under the new classification rules of CHUK03 dataset, the Rank1 value of the proposed method reaches 33.5%. The results show that the accuracy of pedestrian re-identification after distance-weighted fusion is significantly higher than that of the separate feature distance metric, and the effectiveness of the proposed deep region features and algorithm model are proved.

In order to achieve robust, accurate and real-time recognition of surface scratches under complex texture background with uneven brightness, a surface scratch recognition method based on deep neural network was proposed. The deep neural network for surface scratch recognition consisted of a style transfer network and a focus Convolutional Neural Network (CNN). The style transfer network was used to preprocess surface scratches under complex background with uneven brightness. The style transfer networks included a feedforward conversion network and a loss network. Firstly, the style features of uniform brightness template and the perceptual features of the detected image were extracted through the loss network, and the feedforward conversion network was trained offline to obtain the optimal parameter values of network. Then, the images with uniform brightness and uniform style were generated by style transfer network. Finally, the proposed focus convolutional neural network based on focus structure was used to extract and recognize scratch features in the generated image. Taking metal surface with light change as an example, the scratch recognition experiment was carried out. The experimental results show that compared with traditional image processing methods requiring artificial designed features and traditional deep convolutional neural network, the false negative rate of scratch detection is as low as 8.54% with faster convergence speed and smoother convergence curve, and the better detection results can be obtained under different depth models with accuracy increased of about 2%. The style transfer network can retain complete scratch features with the problem of uneven brightness solved, thus improving the accuracy of scratch recognition, while the focus convolutional neural network can achieve robust, accurate and real-time recognition of scratches, which greatly reduces false negative rate and false positive rate of scratches.

Focused on the problem that the memory and computational resource intensive nature of Convolutional Neural Network (CNN) limits the adoption of CNN on embedded devices such as edge computing, a convolutional neural network compression method combining network weight pruning and data quantization for embedded hardware platform data types was proposed. Firstly, according to the weights distribution of each layer of the original CNN, a threshold based pruning method was illustrated to eliminate the weights that have less impact on the network processing accuracy. The redundant information in the network model was removed while the important connections were preserved. Secondly, the required bit-width of the weights and activation functions were analyzed based on the computational characteristics of the embedded platform, and the dynamic fixed-point quantization method was employed to reduce the bit-width of the network model. Finally, the network was fine-tuned to further compress the model size and reduce the computational consumption while ensuring the accuracy of model inference. The experimental results show that this method reduces the network storage space of VGG-19 by over 22 times while reducing the accuracy by only 0.3%, which achieves almost lossless compression. Meanwhile, by evaluating on multiple models, this method can reduce the storage space of the network model by a maximum of 25 times within the range of average accuracy lose of 1.46%, which proves the effective compression of the proposed method.

In order to explore the possibility of using two microphones for separation and locating of multiple sound sources in a two-dimensional plane, an indoor voice separation and sound source localization system based on dual-microphone was proposed. According to the signal collected by microphones, a dual-microphone time delay-attenuation model was established. Then, Degenerte Unmixing Estimation Technique (DUET) algorithm was used to estimate the delay-attenuation parameters of model, and the parameter histogram was drawn. In the speech separation stage, Binary Time-Frequency Masking (BTFM) was established. According to the parameter histogram, binary masking method was combined to separate the mixed speech. In the sound source localization stage, the mathematical equations for determining the location of sound source were obtained by deducing the relationship between the model attenuation parameters and the signal energy ratio. Roomsimove toolbox was used to simulate the indoor acoustic environment. Through Matlab simulation and geometric coordinate calculation, the locating in the two-dimensional plane was completed while separating multiple targets of sound source. The experimental results show that, the locating errors of the proposed system for multiple signals of sound source are less than 2%. Therefore, it contributes to the research and development of small system.

Concerning the problem brought by a large number of spam, an incremental passive attack learning algorithm was proposed. The passive attack learning method was based on the Simple Mail Transfer Protocol (SMTP) session log initiated by the email host in the campus during half a year. Analysis on the status of delivery rate and many types of failure message of the host behavior in the session record was conducted, and the effective adaptation was ultimately achieved by detecting spam source host behavior on the recent email classification. The experimental results show that after implementing several rounds of classification strategy adjustment, the detection accuracy of the proposed model can reach 94.7%. The design is very useful to effectively detect internal spam host and control the spam from the source.

In order to solve tracking failure problem caused by target occlusion, appearance variation and long time tracking in practical monitoring, an object tracking algorithm based on RANdom SAmpling Consensus (RANSAC) estimation was proposed. Firstly, the local invariant feature set in the searching area was extracted. Then the object features were separated from the feature set by using the transfer property of feature matching and non-parametric learning algorithm. At last, the RANSAC estimation of object features was used to track the object location. The algorithm was tested on video data sets with different scenarios and analyzed by using three analysis indicators including accuracy, recall and comprehensive evaluation (F1-Measure). The experimental results show that the proposed method improves target tracking accuracy and overcomes track-drift caused by long time tracking.

In order to solve the problem of information loss caused by local motion blur in daily captured images or videos, a local motion detection algorithm based on region energy estimation was proposed. Firstly, the Harris feature points of the image were calculated, and alternative areas were screened out according to the distribution of feature points of each area. Secondly, according to the characteristic of smooth gradient distribution in monochromatic areas, the gradient distribution of the alternative areas was calculated and the average amplitude threshold was used to filter out most of areas which can be easily misjudged. At last, the blur direction of the alternative areas was estimated according to the energy degeneration feature of motion blur images, and the energy of the blur direction and its perpendicular direction were calculated, thus the monochrome region and defocus blur areas were further removed according to the energy ratio in both above directions. Experimental results on image data sets show that the proposed method can detect the motion blur areas from images with monochromatic areas and defocus blur areas, and effectively improve the robustness and adaptability of local motion blur detection.

Focusing on the mismatch occurred in template matching when using Generalized Hough Transform (GHT) algorithm to extract the target shape from the rotated image, an improved orientation-invariant generalized Hough transform algorithm based on U-chord curvature was proposed. Firstly, the modified R-table with orientation-invariant performance was constructed by using features of U-chord curvature and displacement vectors of edge points of the template shape; secondly, the information such as the displacement vector was achieved by calculating the curvature of edge points as an index to lookup the constructed R-table; finally, the possible locations of reference points were calculated according to the information. The point with maximum voting was the location of the target shape of the image. When the target shape of the image is rotated by 0°, 2°, 4°, 5° and 6° individually, the sharper peaks occur in the target shape position of all the rotation images by using the proposed algorithm. The simulation results show that the Improved Generalized Hough Transform (I-GHT) algorithm has high stability in rotation and noise conditions.

In order to reduce the complexity of signal reconstruction algorithm, and reconstruct the signal with unknown sparsity, a new algorithm named One Projection Subspace Pursuit (OPSP) was proposed. Firstly, the upper and lower bounds of the signal's sparsity were determined based on the restricted isometry property, and the signal's sparsity was set as their integer middle value. Secondly, under the frame of Subspace Pursuit (SP), the projection of the observation onto the support set in each iteration process was removed to decrease the computational complexity of the algorithm. Furthermore, the whole signal's reconstruction rate was used as the index of reconstruction performance. The simulation results show that the proposed algorithm can reconstruct the signals of unknown sparsity with less time and higher reconstruction rate compared with the traditional SP algorithm, and it is effective for signal reconstruction.

Due to the lack of performance analysis while designing a distributed Evolutionary Algorithm (dEA), the designed algorithm cannot reach the expected speedup. To solve this problem, a comprehensive performance analysis method was proposed. According to the components of dEAs, factors that influence the performance of dEAs can be divided into three parts, namely, evolutionary cost, fitness evaluation cost and communication cost. Firstly, the feature of evolutionary cost under different individual encoding lengths was studied. Then when the evolutionary cost was kept unchanged, the fitness evaluation cost was controlled by using the delay function of the operating system and the communication cost was controlled by changing the length of individual encoding. Finally, the effect of each factor was tested through control variable method. The experimental results reveal the constraint relation among the three factors and point out the necessary conditions for speeding up dEAs.

In order to enhance the communication efficiency of distributed heterogeneous system, a clustering-based message-oriented middleware construction technology was proposed, which described the protocols by exploiting the EPr/TN net formally. Accordingly, clustering method was used to deal with the protocols, which obtained the greatest similarity cluster. And then, the protocol clusters were used to construct message-oriented middleware. Compared to the traditional message-oriented middleware, the new technology can not only effectively transform heterogeneous communication protocols, but also reduce the search times of protocol mapping, as well as enhancing the communication efficiency of distributed heterogeneous system.

Static Separation Of Duty (SSOD) is an important principle of information system security. In Role-Based Access Control (RBAC), it is difficult to enforce 2-n SSOD policy directly based on 2-2 Static Mutually Exclusive Role (SMER) constraints. In this paper, the necessary and sufficient conditions of realizing 2-n SSOD policy based on 2-2 SMER constraints were proposed and proved. The sufficient condition proposed was less restrictive than the existing research and allowed more flexible privilege assignment. By the operation rules of authorization management, the sufficient condition was kept and the satisfaction of 2-n SSOD policy during the dynamic change of application environment could be maintained. The application example shows that the method is correct and effective.

In order to reduce the impact of problems inherent in the mutual information method on the filtering effect, a method of candidate term filtration and extraction was proposed. And a determination algorithm based on partial evaluating indicator was given, which can give the best upper and lower thresholds fast and accurately through data sampling, statistics and computation. Compared with the method of mutual information filtration with single threshold, the proposed method filtered and extracted candidate terms by setting two thresholds in the premise of not changing the calculating formula of mutual information. The experimental results show that the proposed method can improve the precision rate and F-measurement significantly under the same conditions.

To the problem that the average recognition rate of person-independent facial expression is not high (about 65%), a new method of facial expression recognition, based on expression subspace and multi-classifiers integration, was proposed. In the training set 1, the features of global face region, eyes (include eyebrows) region and mouth region were respectively extracted and decomposed by Local Binary Pattern (LBP) and Higher Order Singular Value Decomposition (HOSVD), and the corresponding expression subspaces were built. Then the facial images of the training set 2 were trained by Support Vector Machine (SVM) in the expression subspaces and the parameters of fuzzy rule system were conducted. Finally, the expression subspaces and the multi-classifiers ensemble were combined to classify the expressions in test set. The experiments were conducted on JAFFE database and the average recognition rate was 71.43%. The experimental results show that the proposed method effectively reduces the influence caused by facial shape feature and facial expression manner, and it has better recognition rate.

Problems such as high delay, high packet loss rate, low stability and reliability exist in current handoff scheme. To solve these problems, a dual-link communication mechanism and a dual-link selection were proposed, and the data transmission algorithm was presented in this paper to acquire accurate signal quality by smooth processing and control the handoff between two communication links at appropriate time by threshold of difference value and packet forwarding using dual-thread. The experimental results show that compared with the single link mechanism, no delayed pulse exists in dual-link mechanism, the packet loss rate is close to zero and the average throughput is increased by 20%. This mechanism can be applied to many environments owing high-speed mobile subnet such as metro, highway, etc.

When light penetrates through sheet, the perspective images such as images of watermark on the banknotes form. Its histogram has only a single peak. It’s very hard to get good segmentation with old methods. A probability model of distribution for perspective image’s background and objects was put forward, and a formula was deduced to compute the optimized segmentation threshold based on the probability model. The experiment results of show that this new approach can get wonderful segmentation for perspective image, and its computing speed is faster than that of other methods.