Research On Interactive Video Object Segmentation Based On Deep Learning

Visual image perception technology is the main means and way for humans to obtain information in recent years.In recent years,with the further development of modern computer video image technology and the increasingly diversification of its applications,people are particularly concerned with images,especially continuous high-level video The increasing demand for image sequences has also fully aroused the enthusiasm and interest in the continuous exploration and research of video image processing and segmentation technology in recent years.Video image segmentation extracts the targets in a sequence of images based on relatively low-level visual information.Its own follow-up work about the processing of high-level image is based on the necessary theory and rationale provieded by VOS.Such the processing of high-level image includes the recognition of target imaging,the tracking of target and the understanding of image.In the current deep learning field,video target segmentation is mainly divided into semi-supervised segmentation and unsupervised segmentation.For semi-supervised video object segmentation,the ground-truth of the first frame object area is usually manually marked by humans,but manual labeling is time-consuming and laborious,and cannot be unified.In unsupervised video object segmentation,object saliency is a subjective concept,and there are certain ambiguities between different people.Therefore,this paper proposes an interactive video object segmentation technology based on deep learning.The feature of the interactive segmentation algorithm is that it greatly reduces the workload of image segmentation annotation and saves a lot of time.Manually mark the frame of the object of interest in a frame in the video sequence,use the improved graph cut algorithm to obtain the initial segmentation result,and then use the deep learning-based video object segmentation algorithm to pass frame by frame to other video images in the video.Object segmentation to get the object area on all frame images.The user checks the segmentation results,gives new interactive information on the video frames with poor segmentation results,and uses the graph cut algorithm to improve the poor frames again.Finally,re-use the network to improve the segmentation results of other frames until the segmentation results of the entire video sequence are satisfactory.Then this paper analyzes the shortcomings of the proposed segmentation network,and proposes an improved method based on the network structure.Firstly,we use dilated convolution to increase the receptive field on the basic network.Secondly,we introduce attention mechanism on the network,which enables the network to quickly select high-value information from a large number of information in the decoding stage,so as to improve the segmentation accuracy.Finally,the interactive video segmentation algorithm based on deep learning proposed in this paper chooses Davis dataset for segmentation experiment,and compares with some mainstream segmentation algorithms at present.At present,from the final experimental results and data analysis,we can draw a conclusion that the segmentation algorithm in this paper is obviously superior to most of the major methods in the precision of video segmentation.