Research On Object Segmentation In Videos Combining Prior Information And Mixture Models

In recent years,with the development of science and technology,a steady stream of video data has appeared on various network platforms.Faced with a huge amount of data,processing and analyzing these video data brings great challenges to video storage and video content analysis in the field of computer vision and pattern recognition.The video object segmentation method in this paper mainly focuses on segmenting the object of interest in the video from the background information.However,how to detect the object of interest in the video and accurately segment the object does have some difficulties.At present,the main difficulties in object segmentation tasks in video are:(1)the video data has the characteristics of high latitude and high information content;(2)the background information in the video data is complex,and the boundary between the object and the background is not clear;(3)The objects in the video data are easily deformed.Aiming at the characteristics of the video data and the difficulties in the segmentation algorithm,in this paper,we propose two completely different object segmentation algorithms in video that combine prior information with mixture models.The main contributions and innovations of this paper are as follows:(1)A video object segmentation method based on saliency detection and mixture models.In order to complete the task of automatically detecting objects from video data and completing accurate segmentation,this paper proposes an unsupervised video object segmentation method.This method first uses a saliency detection method,which combines the moving edges generated by the object movement and the appearance edges of the object to initialize the object area.The object area is used as a priori information,model the object and background separately using a Gaussian mixture model.The Markov model accurately segmented the object to obtain a pixel-level segmentation map.Experimental results show that the proposed method can quickly and accurately segment the objects in the video.(2)A video object segmentation method based on deep learning and mixturemodels.Only using low-level feature information,good segmentation results cannot be obtained in complex scenes.In order to learn the high-level semantic features in video data,this paper uses the convolutional siamese network to predict the size and position of the object as prior information,and uses the Dirichlet mixture model with spatial constraints to optimize the segmentation results.This method realizes that given the object size and position of the initial frame of a video sequence,the pixel-level annotation of the next frame image is quickly given.Experimental results show that Dirichlet mixture model combining prior information improves segmentation accuracy.The two methods of object segmentation proposed in this paper are verified on the public video data sets Seg Track and Seg Track v2,and compared with other object segmentation methods.The experimental results show that the both methods have good performance.