A Review of Semantic Segmentation Based on Context Information

Abstract

Semantic image segmentation, which is one of the basic tasks in computer vision, is the basis of instance segmentation. It predicts the dense labeling of each pixel in an image and is a very important task that helps us to understand the application scenario. This paper mainly studies the fully convolutional network (FCN) of end-to-end training in semantic segmentation. At the same time, it analyzes the problem that FCN can not handle the context well and the segmentation accuracy is low. There summarize the classical methods which are Conditional Random Fields(CRF), dilated convolutions, Multi-scale feature fusion and deconvolution networks, analyze their performance of semantic segmentation algorithm and explore the possible research directions in the future.