Study On Instance Segmentation Algorithms Based On Mask R-CNN

Instance segmentation is one of computer vision tasks,it is closest to human real visual perception,and it is widely used in the field of automatic driving.Instance segmentation undertakes multiple visual tasks,it not only needs to complete the task of classification and location in object detection,but also needs to segment each different object.Therefore,it is particularly difficult to implement an instance segmentation algorithm with high accuracy.With the introduction of Mask R-CNN,a network framework with good effect and strong expansibility has come into the field of instance segmentation.However,Mask R-CNN is not meticulous enough in dealing with details such as edge segmentation,and has low speed.In order to improve the performance of instance segmentation network,an improved instance segmentation algorithm is proposed in this paper.And a video instance segmentation algorithm is proposed to extend image instance segmentation to video instance segmentation.In this paper,an instance segmentation algorithm based on dilated convolution and edge information is proposed.To avoid information loss caused by pooling and deconvolution,the Hybrid Dilated Convolution is used to enlarge the receptive field and keep the same resolution instead of some pooling operations.In addition,the branch of edge detection is added to the end of the network,and the result of edge detection is regarded as one of the network losses,which improves the accuracy of edge detection and accelerates the convergence of the network.The experimental results show that the average precision of the proposed algorithm is 6.8% higher than that of the original algorithm,and it has a certain competitiveness in the state-of-art algorithms.On the other hand,a video instance segmentation algorithm considering the inter-frame relationship is proposed in this paper.Integrating the deep feature flownet into the network,only key frames is carried out feature extraction.By calculating the flow field and scale field between the key frame and the current frame,the feature map of the current frame is obtained.In order to realize feature propagation among video frames,a method of combining inter-frame relations is proposed,so multilayer features obtained from feature pyramids can be propagated between frames.In this method,feature propagation is adopted instead of feature extraction to improve the running speed of the algorithm.The experimental results show that the proposed algorithm can improve the frame rate of video segmentation while guaranteeing high quality segmentation results.