Semantic Image Segmentation based on SegNetWithCRFs

Abstract

In order to solve the problems of rough segmentation results and loss of image details due to the lack of smooth- ing constraints and continuous downsampling in semantic segmentation tasks. In this article, we propose a end-to-end network model based on SegNetWithCRFs. Conditional Random Fields(CRFs) with Gaussian pairwise potentials and mean-field approximate inference is the last layer of the SegNet network, so that the model has the characteristics of both Deep Convolutional Neural Networks(DCNN) and CRFs, and can learn the parameters of DCNN and CRFs together in a unified deep network. We train the entire deep neural network end-to-end through the backpropagation algorithm, avoiding separate post-processing of the image. Through the qualitative analysis of the experiment on the zebra crossing and KITTI-Road datasets, it can be found that the SegNet model after adding CRFs unified training can effectively solve the problem of image detail loss, even unmarked details can be identified. The algorithm is also effective for segmentation targets with obvious geometric features.