An Instance Segmentation Algorithm Based On Signed Distance Field

As a high-level visual task, instance segmentation aims to distinguish foreground from an image and achieve pixel-level object segmentation. With the rapid development of instance segmentation in recent years, the rich image information extracted from the task of instance segmentation can be widely applied in various fields of work and life. It is a key technology in medical image processing, unmanned driving and other application scenarios. Aiming at the high requirement of segmentation accuracy in application scenarios, an improved instance segmentation algorithm based on signed distance field is proposed in this paper for the first time. Based on the Mask R-CNN, the idea of scale arrangement and cross-scale connection is introduced into the feature extraction network to replace the ResNet + FPN, which further improves the positioning accuracy in the object detection stage. The boundary of instances in the instance segmentation task is always closed, therefore, the signed distance field branch proposed in this paper is introduced after mask branch, which can be useful for predicting the boundary contour of the instance more accurately and solve the problem of rough boundary segmentation in the segmentation task. Experimental results show that the proposed method has better segmentation effect on COCO2017 data set than the benchmark method.