Research On Visual Attention Guided Video Object Segmentation

As video data acquisition and Internet communication become more and more convenient,huge video data need to be processed every day.Efficient video object segmentation technology is one of the key foundations for automatic analysis of massive video data.Video object segmentation is equivalent to video object tracking at the pixel level.According to the degree of human participation,it can be divided into unsupervised segmentation,interactive segmentation and semi-supervised segmentation.Among them,unsupervised segmentation automatically detects the salient objects in video,semi-supervised segmentation requires users to annotate the precise template of the object of interest,and interactive segmentation in between.Supervised segmentation has prominent application prospects,so the deep learning method as the representative of semi-supervised video object segmentation has achieved good research results.However,most methods need to fine-tune network by object template when processing each video sequence,so that the object instance can be adapted well.This online fine-tuning method is very time-consuming and affects its practical application.Aiming at this problem,the paper focuses on the semi-supervised video object segmentation method without online fine-tuning.The main work of this paper is as follows:Firstly,the current classic video object segmentation methods are summarized in detail,including traditional machine learning and deep learning based video object segmentation methods,and the related technologies involved in these methods.Based on this,the existing problems of video object segmentation methods are pointed out,and the purpose of this paper is put forward.Then,based on visual attention mechanism,a guided segmentation method for video object is proposed.Based on the image segmentation network,the visual and spatial information of the interested object is fused as guidance information,and the visual attention is combined to guide the image segmentation network to pay attention to the object,which avoids the process of online fine-tuning and achieves the segmentation of video objects.Thirdly,combining the RNN,the model structure of the above object-guided method is optimized.On the basis of the above object-guided method,the structure of RNN is added to realize online updating of guidance information,improve the object segmentation effect of long sequence video,and realize end-to-end learning of model.Finally,based on the DAVIS dataset,a series of comparative experiments with other methods on the three evaluation indicators of regional similarity,contour accuracy and running speed prove that the model can achieve better without relying on online fine tuning.It shows that this model has certain reference and application value.