Instance Segmentation Problem And Algorithm Research For Real-time Detection Applications

Instance segmentation is a computer vision task that closely resembles human visual perception.It aims to simultaneously address object localization and segmentation.With the advancement of deep learning techniques,numerous algorithms have emerged in the fields of object detection and semantic segmentation,greatly promoting the development of instance segmentation.Based on this foundation,many excellent instance segmenta-tion algorithms have been proposed.Instance segmentation is widely applied in areas such as autonomous driving,robotics systems,intelligent manufacturing,and intelligent transportation.For example,in the track segmentation task of autonomous driving sys-tems in rail transportation,existing methods for detecting and segmenting tracks utilize instance segmentation algorithms based on two-stage detectors,such as Mask-RCNN,to detect and segment the track position^[1].Although this method accurately detects and seg-ments the track position,it falls far short of the real-time application requirements in terms of inference speed.To address the issue of two-stage detector-based instance segmentation algorithms failing to meet the real-time requirements of track segmentation tasks,this the-sis proposes the use of instance segmentation algorithms based on single-stage detectors to rapidly and accurately detect and segment the track position.Additionally,this thesis improves the performance of instance mask segmentation accuracy based on single-stage detectors.The main research content of this thesis is as follows:Firstly,this thesis addresses the issue of inaccurate segmentation results caused by erroneous coefficient prediction during the prototype mask fusion process in instance segmentation algorithms.Based on the Yolcat algorithm,an improved algorithm called Group Yolact is proposed,which not only guarantees inference speed but also enhances the segmentation accuracy of the algorithm.In order to maintain the fast and simple character-istics of the instance segmentation algorithm,the prototype masks are typically generated through linear combination during the prototype fusion process.However,this approach heavily relies on the coefficients for mask fusion.Therefore,this thesis proposes a method of prototype grouping followed by fusion,which to some extent improves the detection and segmentation accuracy of instance masks.Secondly,to address the issue of insufficient attention to instance edge information in instance segmentation algorithms,resulting in less smooth segmentation mask contours,this thesis proposes the Group Yolact Edge algorithm that incorporates edge information to optimize the contours of instance segmentation masks.This method utilizes traditional edge detection operators to extract the edge information of instance masks,and two differ-ent edge detection operators are compared and analyzed in this thesis.It is worth noting that the fusion of mask information only occurs during the process of network training to optimize learnable parameters and does not significantly impact the inference speed of the network.By incorporating edge information,the algorithm improves the optimization of instance edge contours while enhancing the detection and segmentation accuracy of instance masks.Lastly,this thesis applies the improved Group Yolact Edge algorithm to track segmen-tation tasks,achieving fast and accurate segmentation of track regions.In the final track segmentation results,it is observed that the segmentation edges of the track mask exhibit a jagged phenomenon.This thesis analyzes the cause of this phenomenon and proposes a simple optimization strategy to address it.