A Hybrid Model Based on Inception Network and Conditional Random Fields for SAR Image Segmentation

Abstract

Convolutional neural network (CNN) has been acknowledged as an effective tool in the image process. However, CNN with the single-scale convolutional kernels has limited ability to segment synthetic aperture radar (SAR) image with complex scenes. To overcome the problem, a hybrid model, named Inception-CRF model, is proposed, which composes of Inception network and conditional random fields (CRF) model. Firstly, Inception network enlarges the width of the network by setting several parallel convolution operations. Thus, the multi-scale features, which contain various representations and are beneficial for preserving the structure information, are extracted for constructing the pixel-wise unary potential. Secondly, to make full use of the context information, the spatial correlations are constructed as pairwise potential, which is robust against speckle. Finally, the spatial correlations and multi-scale features are incorporated into the posterior distribution to capture information more comprehensively. Experiments on the simulated and real images demonstrate the effectiveness of Inception-CRF model in SAR image segmentation.