Research On Clothes Recognition Method Based On Video Content

With the development of the Internet,people's lives become more colorful.Watch online video has become an important part of leisure life,while the online video advertising has become a major profit source of video company.The stars in the video are leading the fashion trend.People might interest in the clothes of a star in the meantime they watch a video plot.Therefore,the clothes identification based on the video content has a great value.The traditional online video advertising is not based on video content,which not only affect the audience viewing experience,but also fail to mine the content of the video.In recent years,deep learning is widely used in various fields of computer vision,which makes it possible to recognize clothes and achieve exact advertising based video contents.In order to solve the above issues,this thesis proposes a method to identify clothes based on video content.The method proposed in this thesis includes two parts: clothes detection and clothes retrieval.In order to reduce the impact of video ads on the user viewing experience and improve the accuracy of directional advertising,the proposed solution connects the video content with the advertising content of the video company.Since the subject is involved with several field of computer vision,related knowledge will be reviewed on deep learning,object detection and image retrieval etc.In the thesis,we will use deep learning methods to solve the tasks of human detection,pose identification,face verification,and clothes detection.In the process,we compare the performance of different networks in different tasks and the integration of the main target detection framework to complete the task of clothes detection.In the part of clothes retrieval,a fast retrieval network based on hash binary coding is adopted to solve the problem of clothes retrieval.Try to use a variety of network characteristics to solve the problem of precision.In order to evaluate the performance of clothing retrieval,we established two evaluation datasets: one is collected the searching results with a crawler in Baidu image searching engine by using crawler technology,and the other is constructed by using clothes pictures in DailyMail dataset.At last,we use the classic evaluation index of information retrieval to evaluate the performance of the two representation generated by two type of deep neural networks on both datasets.