Semantic Segmentation of PolSAR Images Using Conditional Random Field Model Based on Deep Features

Abstract

Aiming at the problem that the representation ability of traditional features is weakly, this paper proposes a semantic segmentation method based on deep convolutional neural network and conditional random field. The pre-trained VGG-Net-16 model is used to extract more powerful image features, and then the semantic segmentation of images is achieved through the efficient use of multiple features and context information by conditional random fields. The experimental results show that compared with the three methods using traditional classical features, the method achieves the highest overall classification accuracy and Kappa coefficient, indicating that VGG-Net-16 can extract more effective features.