An Instance Segmentation Algorithm Based on Improved Mask R-CNN

Abstract

Mask R-CNN has a high application value in the field of computer vision. However, the Mask R-CNN algorithm has the disadvantages of poor edge segmentation due to the blurred bounding box of the target image and poor segmentation of small targets, which greatly limits its wide application. To solve the above problems, this paper proposes a multi-scale RPN(Region Proposal Network) network structure and adopts KL loss. By building the Tensorflow deep learning framework in the Ubuntu16.04 operating system, the improved algorithm was tested in the MS-COCO data set and the autonomous driving data set Cityscapes which verifies its applicability and effectiveness.