Research On Single-stage Instance Segmentation Method Based On Anchor-free Bo

Instance segmentation belongs to a comprehensive task in the field of computer vision,and its applications involve scenarios such as autonomous driving,intelligent transportation,intelligent security,and intelligent medicare.The current mainstream two-stage instance segmentation methods rely on complex network structures,have cumbersome post-processing steps,and the segmentation speed is far from meeting the needs of application scenarios.The single-stage instance segmentation based on anchor-free correspond to the object and the mask one by one through location information,which is free from the limitation of anchor,but there is still room for further optimization in segmentation accuracy.This thesis focuses on improving the shortcomings and deficiencies of anchor-free single-stage instance segmentation tasks.The main contributions are summarized as follows.For the problems of low discriminability of object features at different scales during feature extraction and difficulty in distinguishing positive and negative samples during training in anchor-free single-stage instance segmentation,this thesis proposes a single-stage instance segmentation method based on shape adaptive sample assignment.Firstly,through the adaptive feature fusion module,the distribution weights of different scale features are learned adaptively to avoid feature information conflicts;secondly,the valid instances are distinguished through the shuffle double attention module to improve the instance location sensitivity.Finally,through the adaptive shape sample assignment strategy,more positive samples are assigned for instance shapes to participate in training without increasing the inference time of the instance segmentation model,and then the segmentation accuracy is improved.For the problems of rough instance segmentation mask boundaries and imprecise instance localization,this thesis proposes a single-stage instance segmentation method based on mask boundary guidance.Firstly,the instance boundary information is extracted by the pyramid boundary extraction module;then the boundary information guidance is provided for instance segmentation by means of boundary mask fusion.Subsequently,the boundary refinement module is used to enhance the feature map context information and retain more boundary details.Finally,the boundary loss is used to constrain the instance boundary for reasonable boundary regression to improve the segmentation accuracy of the model.The single-stage instance segmentation methods proposed in this thesis have been experimented and analyzed in large public datasets COCO and Cityscapes to fully validate the effectiveness of the proposed methods.