An instance segmentation model based on improved SOLOv2 and Chan-Vese

The classical instance segmentation model has the problems of obtaining incomplete feature context information and performing rough segmentation edge smoothing refinement processing, which reduces the segmentation accuracy. To solve these problems, we propose a box-supervised instance segmentation model based on the improved SOLOv2 and Chan-Vese level set method. Firstly, the dilated convolution is introduced into the dynamic convolution kernel prediction module of the SOLOv2 model. The improved SOLOv2 mask-supervised model is used to predict the instance mask, which enlarges the sensing field and obtains rich contextual feature information. Secondly, the box projection function is introduced and utilized to map the instance mask to the initial contour of the Chan-Vese model, thus achieving box-supervised instance segmentation. Finally, an improved length regularization term is added to the Chan-Vese functional to make the object contour edges smoother and segment the object contour effectively. The experimental results show that the proposed instance segmentation model obtains 39.4%, 32.6%, and 22.4% of the masked mAP on the three datasets of Pascal VOC, COCO, and Cityscapes, respectively, which verifies that the proposed method has a better performance for image edge segmentation in general scenes.