To solve the problem that the two-tower neural network in the recommendation system is difficult to learn the interaction information between the user side and the item side and the graph connection information， a new algorithm TSN （Two-stage Siamese graph convolutional Neural network recommendation algorithm） was proposed. First， a heterogeneous graph based on user behavior was built. Then， a graph convolutional Siamese network was designed between the two-tower neural networks， so as to achieve information interaction while learning the connection information of the heterogeneous graph. Finally， by designing a special structure of two-stage information sharing mechanism， the neural networks on the user side and the item side could transmit information dynamically and bidirectionally during the training process， and neural network cascading was effectively avoided. In comparative experiments on MovieLens and Douban movie datasets， the NDCG@10， NDCG@50， NDCG@100 of the proposed algorithm are 11.39% to 23.98% higher than those of the optimal benchmark algorithm DAT （Dual Augmented Two-tower model for online large-scale recommendation）. The results show that the proposed algorithm can alleviate the problem of lack of information interaction in the two-tower neural network； and significantly improves the recommendation performance compared with the previous algorithms.

For the lack of the representation of global information relationship between feature channels in existing monocular image depth estimation tasks, a monocular image depth estimation method based on SE-ResNeXt (Squeeze-and-Excitation-ResNeXt) was proposed. Firstly, the global information representation ability of the network was improved by modeling the dynamic and non-linear relationship between the feature channels. Then, the feature re-calibration strategy was introduced to adaptively re-calibrate the response of feature channel in order to further improve the feature utilization. Finally, the performance of the method was improved without increasing the complexity of the model by using the ResNeXt structure. Experimental results show that compared to the algorithm without ResNeXt structure, the proposed algorithm can obtain lower error, and has the Root Mean Squared Error (RMSE) 10% lower and the Absolute Relative error (AbsRel) 27% lower.

The idea of hierarchical processing of convolution features in neural networks has a significant effect on saliency object detection. However, when integrating hierarchical features, it is still an open problem how to obtain rich global information, as well as effectively integrate the global information and of the higher-level feature space and low-level detail information. Therefore, a saliency detection algorithm based on a multi-level global information propagation model was proposed. In order to extract rich multi-scale global information, a Multi-scale Global Feature Aggregation Module (MGFAM) was introduced to the higher-level, and feature fusion operation was performed to the global information extracted from multiple levels. In addition, in order to obtain the global information of the high-level feature space and the rich low-level detail information at the same time, the extracted discriminative high-level global semantic information was fused with the lower-level features by means of feature propagation. These operations were able to extract the high-level global semantic information to the greatest extent, and avoid the loss of this information when it was gradually propagated to the lower-level. Experimental results on four datasets including ECSSD,PASCAL-S,SOD,HKU-IS show that compared with the advanced NLDF (Non-Local Deep Features for salient object detection) model, the proposed algorithm has the F-measure (F) value increased by 0.028、0.05、0.035 and 0.013 respectively, the Mean Absolute Error (MAE) decreased by 0.023、0.03、0.023 and 0.007 respectively, and the proposed algorithm was superior to several classical image saliency detection methods in terms of precision, recall, F-measure and MAE.

The traditional Total Variation (TV) minimization algorithm is a classical iterative reconstruction algorithm based on Compressed Sensing (CS), and can accurately reconstruct images from sparse and noisy data. However, the block artifacts may be brought by the algorithm during the reconstruction of image having not obvious piecewise constant feature. Researches show that the use of High Order Total Variation (HOTV) in the image denoising can effectively suppress the block artifacts brought by the TV model. Therefore, a HOTV image reconstruction model and its Chambolle-Pock (CP) solving algorithm were proposed. Specifically, the second order TV norm was constructed by using the second order gradient, then a data fidelity constrained second order TV minimization model was designed, and the corresponding CP algorithm was derived. The Shepp-Logan phantom in wave background, grayscale gradual changing phantom and real CT phantom were used to perform the image reconstruction experiments and qualitative and quantitative analysis under ideal data projection and noisy data projection conditions. The reconstruction results of ideal data projection show that compared to the traditional TV algorithm, the HOTV algorithm can effectively suppress the block artifacts and improve the reconstruction accuracy. The reconstruction results of noisy data projection show that both the traditional TV algorithm and the HOTV algorithm have good denoising effect but the HOTV algorithm is able to protect the image edge information better and has higher anti-noise performance. The HOTV algorithm is a better reconstruction algorithm than the TV algorithm in the reconstruction of image having not obvious piecewise constant feature and obvious grayscale fluctuation feature. The proposed HOTV algorithm can be extended to CT reconstruction under different scanning modes and other imaging modalities.

Concerning the influence of changing light, shape, appearance, as well as occlusion on target tracking, a target tracking algorithm based on Speeded Up Robust Feature (SURF) and Multi-Instance Learning (MIL) was proposed. Firstly, the SURF features of the target and its surrounding image were extracted. Secondly, SURF descriptor was introduced to the MIL as the examples in positive and negative bags. Thirdly, all the extracted SURF features were clustered, and a visual vocabulary was established. Fourthly, a "word document" matrix was establish by calculating the importance of the visual words in bag, and the latent semantic features of the bag was got by Latent Semantic Analysis (LSA). Finally, Support Vector Machine (SVM) was trained with the latent semantic features of the bag, so that MIL problem could be handled in accordance with the supervised learning problem. The experimental results show that the robustness and efficiency of the proposed algorithm under the variation of scale, gesture and appearance, as well as short-term partial occlusion.