Research On Edge Accuracy And Stability Of Object Segmentation

Object segmentation aims to identify the contour of the object,whose quality includes the accuracy of edge segmentation in the static image and the edge stability in the video data.In order to improve the internal consistency between features or improve the edge details,the existing segmentation methods often use fusion between different scales to model global context information or take advantage of additional edge weighting mechanisms.By taking the body and the boundaries as the important reference indicators,this thesis proposes a method of instance segmentation based on dynamic convolution for body and boundary decoupling.At the same time,when the segmentation algorithm is applied to video segmentation,the stability of the edge should be improved as well.In order to alleviate the jitter of segmentation algorithm in video data,this thesis starts the study from the perspective of loss function,and proposes a progressive self-supervised stability loss function.The specific research content is as follows:(1)First of all,DBE-CondInst is proposed on the basis of CondInst.The DBE module is designed to substitute the mask branch structure to achieve the purpose of decoupling the body and edge features.Secondly,the encode-decode module in the DBE module is used to filter out the irrelevant noise,the main body features are encoded,then the optical flow is utilized to generate more continuous body and edge features.Finally,the dynamic convolution algorithm is used to dynamically generate dense segmentation branches and then the body and edge features are merged to generate final results,consequently eliminating the dependence of ROI and feature alignment.On the COCO dataset,our test results outperform CondInst by 0.8?1.1 AP with only extra 2%computation cost under different training strategies and are better than other state-of-the-art methods,proving the effectiveness of the proposed DBE-CondInst algorithm,significantly improving the boundary details and internal continuity.(2)When applying segmentation algorithms to video media,the jitter of the prediction result cannot be avoided.The geometric transformations such as rotation,translation,and zoom are used to model the movement of the target from the perspective of loss.Without using video segmentation datasets and additional annotations,we set up a strong to weak self-supervised disturbance method to learn sequence stability step by step.The experimental results show that after adding the proposed progressive self-supervised stability loss function to train the model,the test results improve the smoothness of continuous video frames and enhance the stability of segmentation on the basis of increased accuracy.(3)Finally,the thesis concludes the work that has been done and then gives the prospects of the future work.