Key Technologies Of Customizable Interactive Video Based On Deep Learning

The customizable interactive video refers to the interactive video that can be customized according to different audiences in the process of video editing and production.Currently,there is a great market demand(commodity shopping guide,etc.)due to its advantages of interactivity and a large amount of information.However,there are still two main problems in the production process of customizable interactive video: First,the frame image deterioration phenomena such as object motion blur,video defocus,object posture singularity,and occlusion often appear in the video.It is difficult for traditional video object detection algorithms to achieve high detection accuracy in such video object detection tasks.Second,the video contains a large amount of object information with complicated structures.The current object information management method lacks effective organization management and semantic association of object information.The Chinese national standard of UCL(Uniform Content Label)can efficiently aggregate the messy object information so that the customizable interactive video producer can effectively organize and manage the object information.To solve the problems above,this thesis applies deep learning techniques to customizable interactive video production and proposes an ODV-DL(Object Detection for Video Based on Deep Learning).In addition,an ILOV?UCL(Information Library of Object in Video with UCL)and an IAAMOV?UCL(Information Aggregation Association Method of Object in Video with UCL)are designed respectively.The main research works are as follows:(1)Aiming at the problem that the traditional video object detection algorithms have low accuracy,a video object detection algorithm based on deep learning,named ODV-DL,is proposed in this thesis.Firstly,a video frame features enhancement module is used to extract and optimize frame features in the video.This module regards the enhancement of video frame features as the encoding and decoding process of the video frame sequence,extracts the rich timing information,and uses timing features to effectively improve the quality of the frame feature.Secondly,a video object relationship extraction module constructed by modeling the object relationship is used to exploit the latent semantic relationship between the objects in the video and recode the feature of the candidate boxes through the object relationship,which enriched the feature representation of the candidate boxes.Finally,the ODV-DL classifies and regresses the encoded features to complete the accurate detection of objects in customizable interactive videos.(2)Aiming at the problem of lack of effective organization,management,and semantic association of object information in the current object information management method,this thesis takes advantage of UCL's ability to standardize the structure of information and designs a video object information library based on UCL,named ILOV?UCL.In addition,in order to update and associate object information in time,a video object information aggregation and association method based on UCL,named IAAMOV?UCL,is proposed.Firstly,the triple loss function is used to solve the shortcomings of video object detection algorithms in identifying fine-grained categories to a certain extent.Subsequently,the results of the fine-grained categories after the recognition are quantitatively aggregated to generate semantical rich video information.Finally,according to the generated video information,the semantic weights between the video object information are calculated to fully exploit the semantic association.(3)A prototype system of customizable interactive video production based on deep learning is implemented.The ODV-DL and the IAAMOV?UCL are validated and analyzed through relevant experiments.The experimental results show that the ODV-DL has higher accuracy than traditional video object detection algorithms,and the IAAMOV?UCL can effectively organize,manage,and associate the object information.