Research On Video Object Segmentation Based On Prior Information And Feature Matching

The development of Internet technology has made the way in which people obtain information become more and more diversified,but it has also resulted in a fair amount of redundant information.Therefore,it becomes more difficult to extract the informa-tion of interest from a large amount of data,especially image and video data.The tech-nology of computer vision allows computer to help people extract useful information from images or videos via simulating the image processing capability of human's eyes.As an important part of computer vision,video object segmentation has strong applied value in video coding,video editing,action recognition,and autonomous driving.The development of video object segmentation show a trend of high segmentation accuracy,high segmentation speed and good universality.There are two main chal-lenges to video object segmentation:(1)the video to be segmented usually has prob-lems such as appearance change significantly,object occlusion,object missing or inter-ference from similar objects,which increase the difficulty of video object segmentation;(2)the segmentation accuracy,segmentation speed and universality of algorithms usu-ally cannot both achieve optimal at the same time,so it is necessary to find a balance between the three.Most of the traditional algorithms cannot take into account all the segmentation difficulties in the video itself,nor can they well balance the three require-ments of high segmentation accuracy,high segmentation speed and good universality.To counter the problems above,this thesis views from the demand of video object seg-mentation,focuses on overcoming the challenges faced by video object segmentation,condenses two key problems:"Design of high-performance feature matching method"and "Full utilization and anti-loss of prior information",and carries out a research on video object segmentation based on prior information and feature matching.The main contributions of this thesis are as follows:(1)Proposed a video object segmentation algorithm based on prior probabil-ity and metric learning.Traditional feature matching based video object segmenta-tion methods usually update the template space with features with high classification confidence.Although the segmentation accuracy of the algorithm is improved,the seg-mentation speed slows down as the size of the template space gradually increases.To solve this problem,this thesis proposes a "Fixed-size template space updating strat-egy",which can not only ensure the addition of new useful information when updating the template space,but also keep the size of the template space fixed.At the same time,this thesis proposes a "Prior probability based metric learning" to introduce the prior probability of features into the calculation of the classification probability of features,which can not only reduce the information loss caused by template feature being directly classified as foreground or background,but also improve the accuracy of calculating the classification probability of features.Experimental results on DAVIS datasets demon-strate that the proposed algorithm reaches the state-of-the-art competitive performance and is more efficient in time consumption.(2)Proposed a video object segmentation algorithm based on prior location and matching decoding.In the traditional feature matching based method,noise fea-tures that are highly similar to object features will seriously affect the accuracy of feature classification.In addition,the process of obtaining segmentation masks from the classi-fication probability of features via bilinear interpolation does not make full use of the in-formation of features,and it is easy to lose the detailed information of images.To solve these problems,this thesis proposes a "Training and testing strategy based on feature flow" and a "Decoding module based on matching scores".The former improves the relevancy between adjacent features and reduces the negative impact of noise features on the classification by introducing the location information of features into the train-ing and testing period,the latter enhances the utilization of image features and image detailed information via the feature enhancement module and the feature refining mod-ule.Experimental results on DAVIS datasets demonstrate that the proposed algorithm can effectively improve the relevancy between adjacent features,reduces the negative impact of noise features,and the segmented mask has more detailed information.(3)Designed a video object segmentation prototype system based on prior in-formation and feature matching.In order to meet the requirements of high segmen-tation accuracy,high segmentation speed and good universality of video object seg-mentation,this thesis integrates all the modules proposed to build a unified video object segmentation prototype system.The prototype system adjusts the parameter of all mod-ules through four training periods to build the segmentation system,and achieves the fast and accurate video object segmentation of video sequence through a test period.Experimental results on DAVIS datasets demonstrate that the designed prototype sys-tem can effectively meet the needs of high segmentation accuracy,high segmentation speed and good universality.