Video Retrieval Algorithm And Its Applications Across Multiple Non-Overlapping Camera Views Based On Salience Features And Transfer Learning

The "content" produced by Internet users is growing rapidly. And media data such as images and videos, which come from different channels, is tightly combined with personal information. This paper studies the the content-based video retrieval technology and proposes improved methods in image semantic segmentation,image feature extraction and image retrieval method.We first introduce the video retrieval method based on keywords and content, and then introduce the importance of image feature extraction in video retrieval, and describe the framework of video retrieval process. In the third chapter, we describe the importance of semantic segmentation in video processing, and introduces several typical semantic segmentation methods. Then we propose a novel framework based on multi-feature fusion and aggregated boosting decision forest for semantic annotationThe 2D and 3D features are first extracted from superpixel blocks and fused into a high dimensional vector. The fusion Process can improve the robustness of the features. Then we propose an algorithm to build a week classifier by using a modified integrated splitting strategy for decision trees. And a Markov random field is then adopted to perform global superpixel block optimization to correct the minor errors and make the boundary for semantic annotation smoother. Finally, a boosting strategy is used to aggregate the different week decision trees into one final strong classification decision tree. The experiment results demonstrate the advantages of the proposed method in terms of classification accuracy and computation efficiency over that of existing semantic segmentation methods. In the fourth chapter, we proposes a novel multi-level important salient feature detection method to formulate the appearance model. First data-adapting convolution filters to obtain the important feature map. Then, the salient feature maps are extracted and fused with the important feature maps to produce the multi-level important salient features. Finally, the color histogram features are aggregated with the important salient features to obtain the representative feature vector. The experimental results validate the efficiency of the proposed feature detection method.In the fifth chapter, we describe the basic concepts of ranking learning algorithm and transfer learning method, and then will be ordered cross-domain ranking learning algorithm is applied to video retrieval. The algorithm overcomes the shortcomings of the existing algorithms using only the source domain data, by using the data of the existing tag information in the source domain and the data in the target domain, the positive sample mean value in the target region is estimated, and a more robust adaptive cross domain ranking model is trained, we make use of all the available data from source and target domains as well as constraints in person retrieval. The experimental results show our method achieves better retrieval performance than existing domain adaptation methods.