Research Of Single Object And Multi-object Video Foreground Extraction Algorithm Based On OSVOS

The OSVOS algorithm is a commonly used video single target foreground extraction algorithm.The number of deep learning sample labels is simplified by this algorithm and a better effect is achieved on video object segmentation.However,a mask label is still needed,and the acquisition efficiency of the real mask label is low.Therefore,the algorithm cannot be directly applied to systems with high real-time requirements.In addition,there are some problems in the OSVOS algorithm,including incomplete object segmentation and inaccurate contour positioning in complex scenes such as illumination changes,occlusion,and similar target interference.Based on the OSVOS algorithm,this paper combines the interactive target selection algorithm to achieve real-time extraction of single target foreground;Then,a parallel contour extraction branch network is added to the video foreground segmentation network to achieve accurate extraction of the single target foreground;Finally,training and testing are conducted using the CDNet2014 dataset and the DAVIS2016 dataset.The main research work includes:1.Single object video foreground extraction based on OSVOS:First,to solve the problem that mask label is obtained in video object segmentation by OSVOS,we design a simple interactive video object extraction algorithm(Inter-OSVOS algorithm).The Inter-OSVOS algorithm is inputted the interactive segmentation result as a label to the segmentation network,and is achieved the fast extraction of the video-specific object while preserving the performance of the OSVOS algorithm.Then,to overcome the limitation of insufficient contour localization,an improved Inter-BS-OSVOS algorithm is proposed.A complementary contour branch network added to the segmentation network which is trained to extract the object contour,and the super-pixel alignment and contour restoration are performed to achieve accurate positioning of the foreground object.Finally,network is trained and tested in the CDNet2014 dataset and DAVIS2016 dataset.The results show that the single object extracted by Inter-OSVOS algorithm and OSVOS algorithm has almost no difference in terms of accuracy.However,Inter-OSVOS algorithm can be obtained in real time and it has more robust to complex scene video.In addition,contour accuracy of Inter-BS-OSVOS algorithms has significantly higher than that of OSVOS and Inter-OSVOS algorithms.2.Multi-object video foreground extraction based on OSVOS.First,to solve the problem incompleteness and low precision of the multi-target foreground extraction,the OSVOS-ReID algorithm is proposed.Based on the Inter-BS-OSVOS algorithm,the pixels are accurate to each instance by introducing the instance semantic segmentation,and the multi-target accurate extraction of the specific scene video is realized by adding the re-identification module to retrieve the lost and newly added targets.Then,the new network is trained and tested in the CDNet2014 dataset and DAVIS2017 dataset.The results show that the OSVOS-RelD algorithm has a significantly improvement on the robustness of the complex scene,the accuracy of the extraction objects and the edge accuracy compared with similar algorithms.