Instance Segmentation Method Based On Low-level Feature And Rotation Detection

Instance segmentation is a very challenging task,it requires not only accurate positioning and classification of foreground objects in image,but also accurate segmentation of target contour from complex background.In the case of complex image content,classic instance segmentation network Mask R-CNN often fails to get detailed segmentation results.There is a certain gap between the edge of segmentation result and actual edge of target,which leads to low segmentation accuracy.In addition,the foreground and background of some scene images are complex and contain more noise,and a single image may contain a large number of targets with large scale difference or densely arranged small targets,which makes the image difficult to detect and segment.In order to solve above problems,the framework of Mask R-CNN and the algorithm of rotation detection are studied.The method of instance segmentation assisted by low-level network features and the method of instance segmentation based on results of rotaion detection are proposed respectively.Firstly,the instance segmentation method assisted by low-level features based on Mask RCNN and its segmentation branch is improved.On the basis of original segmentation branch,low-level network features,as auxiliary information,are added to enhance expression ability of segmentation branch features,so as to improve the segmentation quality of target contour edge.The low-level feature fuse with high-level feature of segmentation branch after feature extraction of object region and convolution processing.Then low-level feature with location boundary information and high-level feature with category semantics could form a complementary relationship.This can improve segmentation quality and obtain more accurate and detailed segmentation results.Experimental results show that proposed method has good robustness and effectiveness when using different feature extraction networks.In order to solve the problem of difficult detection and segmentation of image in some scenes,we propose a method to detect rotating object first,and then segment object based on detection results.Specifically,based on existing single-stage detector,we construct a detector that can detect target from any angle in order to classify and locate target accurately from cluttered background.Then,feature pyramid network and prediction layer of preliminary rotation detector are simplified and improved to enhance the ability of distinguishing foreground and background.Finally,a segmentation network is added after detection network,and object is segmented based on rotation detection results.Experimental results show that the improved and enhanced rotation detector can improve the performance of arbitrary angle object detection significantly,and also verify the effectiveness of instance segmentation method based on rotation detection results.