Research On Instance Segmentation Algorithm Based On Mask R-CNN

Instance segmentation is an emerging computer vision task,which not only requires pixel level segmentation of each object of interest in the image,but also needs to distinguish different individuals in each category.This task has high application value in the fields of automatic driving,intelligent medical treatment,robot control and so on.This paper conducts an in-depth study on the instance segmentation model based on the Mask R-CNN,and propose improved models for instance segmentation.The main contents are as follows:(1)In order to solve the problem of insufficient accuracy at the edges of Mask R-CNN segmentation results and improve the effect of instance segmentation,an instance segmentation model based on multi-feature fusion is proposed.Based on Mask R-CNN,an edge detection branch and a semantic segmentation branch are added,which generate feature maps that focus on edge information and spatial location information,respectively.In addition,when ROI alignment is performed,the ROI is mapped to the corresponding pyramid layer and its adjacent layers at the same time to obtain multiple features.Finally,the above feature maps are fused to generate a new feature with more information for subsequent detection and segmentation tasks,which improves the edge details and accuracy of instance segmentation results.Compared with Mask R-CNN,the Mean Average Precision of detection and segmentation on the COCO dataset is improved by 1.2% and 1.0%,respectively.(2)In order to further enhance the feature extraction capability of the Mask R-CNN,an instance segmentation model based on multi-resolution parallelism and attention mechanism is proposed.This model uses a multi-resolution parallel residual network and attention feature pyramid instead of the Resnet101+FPN network in Mask R-CNN to extract image features.At each down-sampling stage of the residual network structure,the multi-resolution residual network adds a parallel branch with the same resolution as that before the down-sampling,and uses the feature output of each branch as the input of the feature pyramid to keep the feature depth of each layer consistent.And the information interaction module makes full use of the information of each layer,especially the lower layers.The attention feature pyramid adds attention modules to each layer of the feature pyramid,so that each layer can adaptively emphasize the information of its important position(such as the large target position of the high-level feature map)when suppressing the secondary information.This model effectively improves the accuracy of instance segmentation,Compared with Mask R-CNN,the Mean Average Precision of detection and segmentation on the COCO dataset is improved by 1.8% and 1.1%,respectively.