PolSAR image segmentation based on feature extraction and data compression using Weighted Neighborhood Filter Bank and Hidden Markov random field-expectation maximization

Abstract

Despite the good performance of methods such as wavelet transform, Gabor filter bank, and scale-invariant feature transform (SIFT) as the first step of image segmentation, they have been some difficulties in the selection of suitable numbers of scales and directions for them. In this paper, a novel method is presented for Polarimetric Synthetic Aperture Radar (PolSAR) image segmentation, in which there is no need for any parameter initialization. The proposed method, called POLWHE is based on Lossy Minimum Descriptive Length (LMDL), Zero Padding Weighted Neighborhood Filter Bank (ZPWNFB) and Hidden Markov Random Field-Expectation Maximization (HMRF-EM). In the proposed method, first, Simple Linear Iterative Clustering (SLIC) and entropy are used for the adaptive generation of superpixels. Then, these superpixels are merged based on the learned texture and color features and the ZPWNFB edge-map. This procedure results in the image being segmented automatically. Next, the initial labels of the image are assigned by k-means and improved by HMRF-EM and ZPWNFB (KWHE). Finally, the automatic method and KWHE are combined and the final segmentation result is obtained. The proposed method is tested on real PolSAR images. The results prove the superiority of the proposed method as it improves both the performance and the noise resistance with the mean accuracy of 92.30%.