Research On Multi-objective Instance Segmentation Algorithm Based On Deep Learning

With the development of computer vision,the application scenarios of the instance segmentation task are becoming more and more frequent,so the instance segmentation technique has received much attention and research from researchers.However,the task has not yet achieved satisfactory results,and most of the algorithms are not accurate enough in the face of complex environments and still fall short of human eye recognition segmentation.In order to solve the problem of using instance segmentation algorithms with high accuracy in engineering,this thesis improves on the Mask R-CNN algorithm,with the aim of solving the problem that instance segmentation techniques can effectively perform instance segmentation in the face of multiple target occlusions,low illumination images,small targets and other situations,and further improve the accuracy of the algorithm.The main research contents and contributions of this paper are as follows:(1)In this paper,Mask-RCNN based multi-scale double layer decomposition model(MBDN)is designed for the multi-target overlap problem,based on the Mask R-CNN algorithm for improvement.The algorithm adopts the detection head of Mask R-CNN as the basic part of the detection of this paper’s algorithm,and the segmentation part decouples the overlapping objects into two image layers,where the top layer deals with the occluded objects and the bottom layer deals with the target objects,and adds multi-scale expansion convolution and graph convolution in each layer for processing,which can effectively acquire more features of the image and decode the image from the global and local perspectives of interest respectively The algorithm is enhanced to obtain global and local features,thus improving the segmentation performance of the image and effectively solving the problem of segmenting multi-target occlusion instances.On the backbone networks Res Net50 and Res Net101,the model improves the accuracy of the baseline model by 7.3%and 7.2% respectively,and the proposed method is validated in COCO public dataset experiments.The experimental results show that the proposed method outperforms similar existing methods.(2)To address the problem of low-illumination image instance segmentation,this paper designs an instance segmentation method based on the MBDN model for low-illumination image enhancement.The method uses an image enhancement module in detection to increase the brightness of the instance segmentation image and increase its detectability.At the same time,the CBAM attention module is added to the detection head to improve the detection capability of the model for small targets.The experimental results of the model outperformed similar existing methods,with the AP improving from 36.85 to 39.33,an improvement of 6.7%,effectively solving the problem of image instance segmentation in complex environments.In summary,this paper has explored and researched the image instance segmentation problem in complex environments,and proposed and validated two deep learning-based instance segmentation models,which are effective compared with other methods and have certain reference significance.