Image Semantic Segmentation Method Based on Atrous Algorithm and Convolution CRF

For semantic image segmentation, much of the work is done using the conditional random field probability graph model as the backend processing for CNN. Among the more successful work, DCNN introduces atrous convolution to extend the perceptual domain without increasing the number of parameters and the amount of computation; capturing objects and contexts of images from multiple scales through the atrous spatial pyramid pool; by combining DCNN and Probability Graph model to enhance the positioning of object boundaries. However, conditional random fields have the disadvantage of slow training and reasoning and difficulty in learning internal parameters. We try to introduce convolutional CRF, solve the problem of CRF through convolution operation, and can be effectively implemented on the GPU. Our proposed algorithm has achieved good results in the PASCAL VOC-2012 semantic image segmentation task.