Video Object Segmentation Based On Deep Learning

Video object segmentation is a fundamental and key technology in computer vision.Its goal is to segment the areas of interest in the video sequence according to certain criteria,including pedestrians appearing in static scene video,the objects capturing in a dynamic scene video and so on.This technology is widely used in intelligent video surveillance,video conferencing,intelligent transportation and other fields.The existing video object segmentation method can be divided into semi-supervised traditional methods,unsupervised traditional methods,semi-supervised deep learning methods,and unsupervised deep learning methods.The semi-supervised methods require human-computer interaction,in the case,the range of application is very limited.It's difficult to use the unsupervised traditional methods to maintain a good segmentation effect in a complex and variable environment,because their feature extractors are designed by humans.The unsupervised deep learning methods don't require people to design the feature extractors,since features are autonomously learned by the network.Thus,the unsupervised deep learning methods have better adaptability to different scenarios.However,There is still a shortage of the unsupervised deep learning methods.The segmentation obtained directly through the network is unlikely acquire complete object without background region which is misidentified as the object.For purpose of improving the accuracy and make the method applicable to more scenes,this paper proposes a cover-seed fusion network based on unsupervised deep learning.The cover-seed fusion network consists of a cover branch,a seed branch,and a fusion part.The cover branch is used to solve the problem of the segmented target object defect or the boundary inaccuracy,the seed branch and the fusion part solve the problem that the background region is incorrectly classified.The two branches use the same neural network structure,and the fusion part uses the region growing algorithm in the traditional image processing method.The models of the two branches are selected by different loss functions and by different evaluation indicators.The fusion part uses the output of the seed branch as the seed of the region growing algorithm,and uses the output of the cover branch as one of the growth constraints.Finally,the final segmentation result is obtained by growing the output of the seed branch.Further,according to different structures in the branch,the cover-seed fusion network can be divided into a cover-seed fusion network that does not utilize inter-frame information and a cover-seed fusion network that utilizes optical flows.In order to verify the effectiveness of the various parts of the network and the overall performance of the network,In this paper,the elimination experiments of each part of the two coverage-seed fusion networks were carried out.In addition,by comparing with other classical video object segmentation methods on the three public video object segmentation data sets,the experimental results demonstrate the effectiveness of the network.