Aiming at the problems of slow detection and low recognition accuracy of road traffic signs in Chinese intelligent driving assistance system， an improved road traffic sign detection algorithm based on YOLOv3 （You Only Look Once version 3） was proposed. Firstly， MobileNetv2 was introduced into YOLOv3 as the basic feature extraction network to construct an object detection network module MN-YOLOv3 （MobileNetv2-YOLOv3）. And two Down-up links were added to the backbone network of MN-YOLOv3 for feature fusion， thereby reducing the model parameters， and improving the running speed of the detection module as well as information fusion performance of the multi-scale feature maps. Then， according to the shape characteristics of traffic sign objects， K-Means++ algorithm was used to generate the initial cluster center of the anchor， and the DIOU （Distance Intersection Over Union） loss function was introduced to combine DIOU and Non-Maximum Suppression （NMS） for the bounding box regression. Finally， the Region Of Interest （ROI） and the context information were unified by ROI Align and merged to enhance the object feature expression. Experimental results show that the proposed algorithm has better performance， and the mean Average Precision （mAP） of the algorithm on the dataset CSUST （ChangSha University of Science and Technology） Chinese Traffic Sign Detection Benchmark （CCTSDB） can reach 96.20%. Compared with Faster R-CNN （Region Convolutional Neural Network）， YOLOv3 and Cascaded R-CNN detection algorithms， the proposed algorithm has better real-time performance， higher detection accuracy， and is more robustness to various environmental changes.

Aiming at the problems of medical ultrasound images such as many noisy points, fuzzy boundaries, and difficulty in defining the cardiac contours, a new Spatial Frequency Divided Attention Network for ultrasound image segmentation (SFDA-Net) was proposed. Firstly, with the help of Octave convolution, the high and low-frequency parallel processing of image in the entire network was realized to obtain more diverse information. Then, the Convolutional Block Attention Module (CBAM) was added for paying more attention to the effective information when image feature recovered, so as to reduce the loss of segmenting the entire target area. Finally Focal Tversky Loss was considered as the objective function to reduce the weights of simple samples and pay more attention on difficult samples, as well as decrease the errors introduced by pixel misjudgment between the categories. Through multiple sets of comparative experiments,it can be seen that with the parameter number lower than that of the original UNet++, SFDA-Net has the segmentation accuracy increased by 6.2 percentage points, Dice sore risen by 8.76 percentage points, mean Pixel Accuracy (mPA) improved to 84.09%, and mean Intersection Over Union (mIoU) increased to 75.79%. SFDA-Net steadily improves the network performance while reducing parameters, and makes the echocardiographic segmentation more accurate.

In order to improve the precision and speed of existing remote sensing image target detection algorithms in small-scale target detection, a remote sensing image target detection and identification algorithm based on deep learning was proposed. Firstly, a dataset of remote sensing images with different scales was constructed for model training and testing. Secondly, based on the original Single Shot multibox Detector (SSD) network model, the shallow feature fusion module, shallow feature enhancement module and deep feature enhancement module were designed and fused. Finally, the focal loss function was introduced into the training strategy to solve the problem of the imbalance of positive and negative samples in the training process, and the experiment was carried out on the remote sensing image dataset. Experimental results on high-resolution remote sensing image dataset show that the detection mean Average Precision (mAP) of the proposed algorithm achieves 77.95%, which is 3.99 percentage points higher than that of SSD network model, and has the detection speed of 33.8 frame/s. In the extended experiment, the performance of the proposed algorithm is better than that of SSD network model for the detection of fuzzy targets in high-resolution remote sensing images. Experimental results show that the proposed algorithm can effectively improve the precision of remote sensing image target detection.

In order to solve the problem that the traditional cascaded Convolutional Neural Network (CNN) has low accuracy of feature point localization in left ventricular ultrasound image, an improved cascaded CNN with region extracted by Faster Region-based CNN (Faster-RCNN) model was proposed to locate the left ventricular endocardial and epicardial feature points in ultrasound images. Firstly, the traditional cascaded CNN was improved by a structure of two-stage cascaded. In the first stage, an improved convolutional network was used to roughly locate the endocardial and epicardial joint feature points. In the second stage, four improved convolutional networks were used to fine-tune the endocardial feature points and the epicardial feature points separately. After that, the positions of joint contour feature points were output. Secondly, the improved cascaded CNN was merged with target region extraction, which means that the target region containing the left ventricle was extracted by the Faster-RCNN model and then was sent into the improved cascaded CNN. Finally, the left ventricular contour feature points were located from coarse to fine. Experimental results show that compared with the traditional cascaded CNN, the proposed method is much more accurate in left ventricle feature point localization, and its prediction points are closer to the actual values. Under the root mean square error evaluation standard, the accuracy of feature point localization is improved by 32.6 percentage points.

For the assumption of unreasonable conditions in speech recognition by Hidden Markov Model (HMM), the ability of sequence modeling of recurrent neural networks was further studied, an acoustic model based on Bidirectional Long Short-Term Memory (BLSTM) neural networks was proposed. The training criterion based on Connectionist Temporal Classification (CTC) was successfully applied to the acoustic model training, and an end-to-end Chinese speech recognition system was built which does not rely on HMM. Meanwhile, a speech decoding method based on Weighted Finite-State Transducer (WFST) was designed to effectively solve the problem that lexicon and language model are difficult to integrate into the decoding process. Compared with the traditional GMM-HMM system and hybrid DNN-HMM system, the experimental results show that the end-to-end system not only significantly reduces the recognition error rate, but also significantly improves the speech decoding speed, indicating that the proposed acoustic model can effectively enhance the model discrimination and optimize the system structure.

The image segmentation method based on appearance-model has high computational complexity in iterative positioning feature points, and it is difficult to optimize the nonlinear local feature. To solve these above problems and locate feature points of left ventricular endocardium and epicardium, a gradient decent algorithm based on supervised learning was proposed, a multi-resolution pyramid model of 4 levels was built, and a new feature extraction function based on Bhattacharyya coefficient, namely B-SIFT, was used to replace the Scale Invariant Feature Transform (SIFT) feature in the original method. Firstly, the training set images were normalized to unify the size of each TransEsophageal Echocardiography (TEE). Then the supervised descent model based on B-SIFT and multi-resolution pyramid was built to get a gradient descent direction sequence that approaches the actual values. Finally, the learned direction sequence was applied to the test set to obtain the segmentation results of left ventricular. The experimental results show that compared with the traditional gradient decent method based on supervised learning, the average segmentation error of the proposed method is reduced by 47%, and the iteration results are more closer to the actual values compared with the single-scale method.

To deal with the problems that the routing algorithm based on Collaborative communication of Cluster Parent (CRPL) for Low Power and Lossy Network (LLN) can't balance the energy consumption of the node and maximize the extension of the lifetime for network efficiently due to take no account of the residual energy of the node, a high-efficient routing algorithm based on collaborative communications of cluster parent set HE-CRPL was proposed. The proposed algorithm chiefly carried out three optimization schemes. Firstly, the wireless link quality and the residual energy of node could be considered during the cluster parent selection. Secondly, the wireless link quality and the Expected LifeTime (ELT) of cluster parent node were combined while estimating the priority of the cluster parent node and selecting the optimal cluster parent set. Thirdly, the cluster parent nodes were notified the priority list by Destination Advertisement Object (DAO) message during the initialization of the network topology. The simulation results show that, compared with the CRPL algorithm, the performance of the HE-CRPL algorithm is improved obviously in prolonging the network lifetime, increasing the packet delivery success rate and reducing the number of packet retransmissions, and that the lifetime of network prolonging by more than 18.7% and the number of retransmissions decrease by more than 15.9%.

B-mode cardiac ultrasound image segmentation is a fundamental step before cardiac functional parameters estimation. Aiming at the problem that the accuracy of segmentation is low because of the low resolution of ultrasound image, and the model based image segmentation algorithms need a large number of training sets, an image segmentation algorithm based on pixel clustering was proposed combined with prior knowledge of B-mode cardiac ultrasound images. Firstly, anisotropic diffusion was used to preprocess the image. Secondly, one-dimensional K-means was used to cluster the pixels. Finally, every pixel value of the image was assigned to the pixel value of its best cluster center according to cluster results and prior knowledge. The theoretical analysis shows that the proposed algorithm can get the maximum Peak Signal-to-Noise Ratio (PSNR) of ultrasound image; the experimental results show that the proposed algorithm performs better than Otsu algorithm, and its PSNR is increased by 11.5% compared with Otsu algorithm. The proposed algorithm can still work even for a single ultrasound image and can be suitble for ultrasound image segmentation of any shapes, so it is conducive to estimate cardiac functional parameters more accurately.

The transmission energy consumption of nodes does not be considered in the exiting Network Coding Aware Cooperative MAC (NCAC-MAC) protocol for Ad Hoc Network, and the control message sent by the candidate cooperative relay node can not make the other candidate nodes which are not in the communication range give up competition, thus causing collision. To deal with those problems, a high efficiency Medium Access Control (MAC) protocol based on cooperative network coding High Efficiency MAC protocol based on Cooperative Network Coding (HECNC-MAC) was proposed. Three optimization schemes were carried out by the protocol. Firstly, candidate cooperative relay node need to prejudge whether the destionation can decode the packet, so as to reduce the number of competitive relay nodes and ensure that the destination node could be successfully decoded. Secondly, the transmission energy consumption of nodes should be synthetically considered when selecting the cooperative relay node. Finally, the Eager To Help (ETH) message is canceled, and the destination node sents conformation message through pseudo-broadcast. Theoretical analysis and simulation results show that in the comparison experiments with Carrier Sense Multiple Access (CSMA), Phoenix and NCAC-MAC protocols, the transmission energy consumption of nodes and the end-to-end delay of date packages can be effectively reduced, and the network throughput can be improved by HECNC-MAC.

To deal with the problems that message-distributed tasks are backlogged by nodes in the inner community and active nodes are blindly selected to transmit message in the Community-based Message Transmission Scheme in Opportunistic Social Network (OSNCMTS), a High-Efficient Community-based Message Transmission Scheme in opportunistic network (HECMTS) was proposed. In HECMTS, firstly, communities were divided by Extremal Optimization (EO) algorithm and the corresponding community matrices were distributed to nodes; secondly, the copies of message were assigned based on the community matrices and the success rate of data packets to destination nodes; finally, the active nodes' information was collected by the opportunities of active nodes' back and forth in different communities, then the suitable nodes were selected to finish message transmitting between communities by querying the active node information. The simulation results show that routing overhead in HECMTS is decreased by at least 19% and the average end-to-end delay is decreased by at least 16% compared with OSNCMTS.

During the automatic segmentation of cardiac structures in echocardiographic sequences within a cardiac cycle, the contour with weak edges can not be extracted effectively. A new approach combining Speeded Up Robust Feature (SURF) and Chan-Vese model was proposed to resolve this problem. Firstly, the weak boundary of heart chamber in the first frame was marked manually. Then, the SURF points around the boundary were extracted to build Delaunay triangulation. The positions of weak boundaries of subsequent frames were predicted using feature points matching between adjacent frames. The coarse contour was extracted using Chan-Vese model, and the fine contour of object could be acquired by region growing algorithm. The experiment proves that the proposed algorithm can effectively extract the contour of heart chamber with weak edges, and the result is similar to that by manual segmentation.

Feature extraction is a key step of image retrieval and image registration, but the single feature can not express the information of medical images efficiently. To overcome this shortcoming, a new algorithm for medical image retrieval combining global features with local features was proposed based on the characteristics of medical images. First, after studying the medical image retrieving techniques with single feature, a new retrieval method was proposed by considering global feature and relevance feedback. Then to optimize the Scale-Invariant Feature Transform (SIFT) features, an improved SIFT features extraction and matching algorithm was proposed. Finally, in order to ensure the accuracy of the results and improve the retrieval result, local features were used for stepwise refinement. The experimental results on general Digital Radiography (DR) images prove the effectiveness of the proposed algorithm.

Irregular computing exists in large scale parallel application widely and the automatic parallelization on distributed memory is hardly to generate parallel code for irregular loops at compile-time. The communication code of the parallel code influences the correctness and the efficiency to the runout of the program. It could automatically generate useful communication code for a common class of irregular loops at compile-time by using the approach of partial communication redundancy, that needed analyzing the array redistribution graph of the program to maintain the producer-consumer relation of irregular array references. The approach searched the local definition set of the irregular array on each processor by computation decomposition and accessed expression of array references as the communication data set, then analyzed the communication strategies for such irregular loops and generated the corresponding communication code. The experimental results show the validity of the approach and the expectant speedup of test applications.

Medical image filtering process should retain the edge details of diagnostic significance. For Perona-Malik (PM) anisotropic diffusion model experienced failure when dealing with strong noise and choosing parameter K of diffusion threshold relies on experience, this paper proposed an improved anisotropic diffusion algorithm. First, PM was combined with the median filter algorithm, and then the gradient mode of the original image was replaced with the gradient mode from the image which was smoothed by the median filter to control the process of diffusion. While applying the adaptive diffusion threshold (Median Absolute Deviation (MAD) of the gradient in current neighborhood) and iteration termination criteria, the algorithm improved robustness and efficiency of the algorithm. The experiment was operated respectively on echocardiography, CT images and Lena image to denoise, and used Peak Signal-to-Noise Ratio (PSNR) and Edge Preservation Index (EPI) as evaluation criterion. The experimental results show that the improves algorithm outperforms PM algorithm and Catte-PM method for improving PSNR while preserving image detail information, and meets the requirements for application in medical images more effectively.

Concerning the poor quality of recommendations of traditional user-based and item-based collaborative filtering models, a new collaborative filtering model combining users and items predictions was proposed. Firstly, it considered both users and items, and optimized the similarity model with excellent performance dynamically. Secondly, it constructed neighbor sets for the target objects by selecting some similar users and items according to the similarity values, and then obtained the user-based and item-based prediction results respectively based on some prediction functions. Finally, it gained final predictions by using the adaptive balance factor to coordinate both of the prediction results. Comparative experiments were carried out under different evaluation criteria, and the results show that, compared with some typical collaborative filtering models such as RSCF, HCFR and UNCF, the proposed model not only has better performance in prediction accuracy of items, but also does well in the precision and recall of recommendations.

The traditional User-based Collaborative Filtering (UCF) models do not consider the attributes of items fully in the process of measuring the similarity of users. In view of the drawback, this paper proposed two collaborative filtering recommendation models considering item attributes. Firstly, the models optimized the rating-based similarity between users, and then summed the rating numbers of different items by users according to item attributes, in order to obtain the optimized and attribute-based similarity between users. Finally, the models coordinated the two types of similarity measurements by a self-adaptive balance factor, to complete the item prediction and recommendation process. The experimental results demonstrate that the newly proposed models not only have reasonable time costs in different data sets, but also yield excellent improvements in prediction accuracy of ratings, involving an average improvement of 5%, which confirms that the models are efficient in improving the accuracy of user similarity measurements.

The existing parallel cost models are mostly devised for shared memory or distributed memory architecture, thus not suitable for heterogeneous multi-core processors. In order to solve the problem, a new parallel cost model for heterogeneous multi-cores was proposed. It described the impact of computing capacity, memory access delay and data transfer cost on parallel execution time of loops quantitatively, thus improving the veracity of accelerated parallel loop recognition. The experimental results show that the proposed model can effectively recognize the accelerated parallel loops. Using its recognition results to generate parallel codes can improve the performance of parallel programs on heterogeneous multi-core processors significantly.

Today, SIMD (Single Instruction Multiple Data) technology has been widely used in Digital Signal Processor (DSP), and most of the existing compilers realize automatic vectorization functions. However，the compiler cannot support SIMD auto-vectorization with the feature of DSP, because of DSP complex instruction set, the specific addressing model, the obstacle of dependence relation to vectorization non-aligned data or other reasons. In order to solve this problem, in this paper, for the automatic vectorization in the Superword Level Parallelism (SLP) based on the Open64 compiler back end, the instruction analysis and redundancy optimization algorithm were improved, so as to transform more efficient vectorized source program. The experimental results show that the proposed method can improve DSP performances and reduce power consumption efficiently.

To resolve the problem that underground pipe network of plant is large, complex and difficult to manage, the C/S+B/S mixed development model based on the ArcGIS platform was used. This paper aimed to transfer the spatial analysis of the traditional Geographic Information System (GIS) from C/S to B/S. Collision detection algorithm and efficient linear interpolation method were used in B/S client, and moreover an optimized relational database management model was proposed. The system running results show that the system could make spatial analysis of the invisible underground pipe network better and more effective, and achieve the digital management of underground pipe network in large plant.

According to the characteristics of wireless single-hop broadcast networks, a network coding scheme based on sliding windows named NCBSW was proposed. The scheme designed a coding window which slid in a chronological order in the matrix of data packets waiting for retransmission, and the data packets used to encode were chosen from the sliding window. Meanwhile, the scheme ensured the solvability of coded packets. The simulation results show that the proposed scheme has a better performance as compared to the retransmission approach in wireless broadcasting based on network coding (NCWBR) in terms of the number of retransmission, delay, network overhead and energy consumption.

Based on the original relative gradient operator, the authors proposed a new relative gradient operator and combined it with 2DPCA or 2DFLD. A face recognition algorithm based on this new relative gradient operator was also put forward. Experimental results on AR and Yale_B face database show that, the method has robustness to the complex change like various expression and lighting condition. The recognition accuracy of the method is much higher than 2DPCA, 2DFLD and face recognition based on the original relative gradient operator. What is more, experiments were done on three different sizes of windows, which confirms that, when the window size is 3×3, the recognition result is relatively the best.