Research On Clothing Retrieval Technology For Video Scene

With the rapid development of Internet e-commerce, online shopping with its convenient characteristic is gradually becoming a new choice of shopping for the majority of users. While the search engine is the most important entrance of the online shopping system. In order to meet this kind of demand, the search technologies of shopping based on the carrier of images begin to attract people's attention. In this paper, the image retrieval technologies were introduced and applied to the video-oriented application scene. Finally, both design and implementation of the shopping search algorithms were completed. The main contents and contributions of this thesis are as follows:(1) This paper introduced and analyzed the important technologies of feature extraction and classification in the clothing retrieval process.Based on the combination of the practical application scenarios and person re-identification algorithm, a video segmentation method was innovatively proposed as a video stream preprocessing step, breaking the limitation to camera locations and providing the premise for the subsequent key frame extraction and multiple frame voting.(2) A more effective fusion feature description was designed. In addition to extracting the statistical histogram-like features, including color, edge and texture, the covariance-expectation feature was also introduced to describe the correlation between features in a specific region, improving the recognition accuracy of the entire clothing retrieval process.(3) The feature fusion algorithm was optimized. Based on the traditional kernel-based classification model, a low-dimensional feature mapping function was introduced to approximate the kernel function,which simplified the mapping process of the kernel function. Not only ensuring the classification performance, but also avoiding the high computational complexity and storage consumption during the kernel training and testing process. In addition, the paper also explored the feature extraction based on deep learning network, and compared with the proposed method, opening up the new horizons for the further research direction.This paper focused on the research of key technologies of clothing retrieval process for the video scene, and implemented design and development of the key algorithms, which has certain significance to the application development for the shopping search technologies.