Research On Instance Segmentation Technology Based On Deep Learning

With the advent of the information age,the image segmentation algorithm based on deep learning has become a hot topic for researchers,and has been effectively applied in many aspects of life and production.As an important field in image segmentation,instance segmentation has become a very popular research direction with its practicability and scalability.It has a large application space in the fields of automatic driving,video monitoring and safety supervision.Instance segmentation not only classifies pixels,but also distinguishes different instances.The Mask R-CNN algorithm is one of the classic instance segmentation algorithms.This paper focuses on the Mask R-CNN algorithm for instance segmentation,understanding of the Mask R-CNN model,and it is improved and optimized,specific work is done as follows:(1)Each module of Mask R-CNN model and the working mechanism of these modules are deeply analyzed,and improves the feature fusion process of the feature pyramid based on the Mask R-CNN algorithm.The original top-down path is changed to bottom-up to retain more spatial position information in the shallow feature map,and dilated convolution is added in the feature fusion process.Dilated convolution can enlarge the receptive field by introducing dilation-rate,while keeping the size of the feature map unchanged.In order to get the receptive fields of different scales,three different dilation-rate are set in this experiment to ensure the acquisition of multi-scale feature information.(2)The modified model be used for the human segmentation dataset,and the human instance segmentation is taken as the application example.The theory is combined with the demand of human instance segmentation.According to the length width ratio of human in practical application,the length width ratio is adjusted.And the human instance segmentation dataset is made for the experiment of the improved human instance segmentation algorithm to verify the feasibility of the algorithm.The outcome make known that the improved Mask R-CNN can achieve better segmentation effect on human segmentation dataset.