Research On Boundary Precision Of Segmentation Based On Deep Learning Case

With the development of artificial intelligence,probabilistic graph model and machine learning,computer vision has made a major breakthrough,and image instance segmentation has become a new challenge in computer vision after semantic segmentation and target detection.Compared with traditional computer vision tasks,instance segmentation requires not only distinguishing different objects in the image but also distinguishing different individual cases among objects,which is highly integrated.At present,there are many algorithms to solve the instance segmentation task,among which,the Mask Region Convolutional Neural Networks(Mask R-CNN)algorithm is relatively mature,but there is still room for improvement in the Mask R-CNN algorithm,especially the large error in the segmentation boundary.In order to solve the Mask R-CNN boundary problem,this article made an in-depth study and proposed an improved scheme.The main content of this article is as follows.First of all,this article summarizes the research topic and background,explains the relevant principles of case segmentation,and expounds the research progress of case segmentation at home and abroad in recent years,and makes a detailed introduction to the classical segmentation model.Secondly,in order to solve the fuzzy problem of boundary segmentation,this article studies the Mask R-CNN and conditional random field,the method in use convolution Mask branch conditions with the airport and all connection with the airport to optimize Mask branch for further segmentation candidate area,makes the result more accurate boundary,and verified by experiment.Thirdly,in order to solve the inaccuracy of boundary segmentation caused by the instance segmentation border regression,this article studies the improved recommendation layer,so that boundary regression can retain a part of background pixels to reduce the loss of pixel information,and prove its effectiveness through experiments.Finally,aiming at the unbalanced problem of instance segmentation in data selection,feature graph,loss function,etc.,this article uses balanced training to improve the final model,so that the training results converge in a more precise direction,and also proposes a fuzzy image training method to further improve the model Stability,and design experiments prove its feasibility.