Research On Instance Segmentation Based On Convolutional Neural Network

Computer vision technology is now widely used in various application fields and industries.Through the use of advanced computer vision technology,computers can perform tasks requiring human vision participation,such as autopilot,robotics and other fields.Therefore,computer vision technology has high research value and practical application significance in academia and industry.Instance segmentation of image processing plays an important role in many visual tasks.It not only needs to complete the same instance classification and instance positioning regression as object detection,but also needs to make pixel annotation for different instances.Therefore,it is very difficult to implement a high precision and speed instance segmentation algorithm.Based on Center Net,this paper proposes a series of instance segmentation methods for operational efficiency,aiming to improve the speed and precision of instance segmentation task,and obtain competitive results in the authoritative data set MS COCO.Firstly,the one-stage instance segmentation algorithm is implemented based on Center Net,and the results of instance segmentation are obtained once by calculating the matching relationship between instance and pixel by embedding.By learning the representation of instances and pixels in highdimensional space,this paper proposes the mask loss with learnable margin to optimize the embedding distance between instances and pixels,and then allocates pixels to corresponding instances in the way of metric learning to generate the segmentation results of instances.Through the MS COCO Instance Segmentation Challenge,it is proved that the precision of the algorithm is comparable to the mainstream algorithm,and it can have a high inference speed.Then,in order to further improve the performance of the instance segmentation algorithm,an independent edge subnet is designed by using the gated convolutional layer to process the edge information of the input samples and fuse it implicitly into the feature graph for prediction.The edges of the prediction mask are extracted by the Sobel operator,and the edge prediction of the segmentation mask is supervised explicitly.In the MS COCO Instance Segmentation Challenge,it achieved 39.2 AP,which was better than the mainstream instance segmentation algorithm.Finally,the computational redundancy of the pixel level segmentation method is analyzed,and the modeling of the instance mask by the polygon contour is proposed to eliminate the redundancy,so as to improve the efficiency of the algorithm.This topic designed the definition of example polygon contour,by predicting the polygon vertices in polar coordinates,connecting the vertices to get the object contour,so as to generate the mask of the object.In the MS COCO Instance Segmentation Challenge,38.6 AP was achieved at 17.4 FPS,achieving both speed and performance improvements.