Novel class-relativity non-local means with principal component analysis for multitemporal SAR image change detection

Abstract

ABSTRACTOne of the main problems for change detection in multitemporal synthetic aperture radar (SAR) images is the presence of speckle noise, since it degrades the image quality significantly and may hide important details in the image. In this article, we investigate a novel class-relativity non-local means (CRNLM) algorithm that reduces the effect of speckle noise in the principal component analysis (PCA) feature space for SAR image change detection. Note that the non-local means averaging process is particularly true when the assumed noise model is additive. Thus, we adopt the difference image produced by the ratio image expressed in logarithmic scale and then transform it onto PCA space. This is done so that its signal energy is concentrated, and the noise spreads over the whole PCA space and is additive. A task-dependent CRNLM algorithm is applied to the PCA transformed data set so as to combine local and non-local geometries and capture the robustness to noise. The idea is based on the assumption that non-local similar patches have similar class structures. Visual and quantitative results obtained on real multitemporal SAR image data sets confirm the effectiveness of this method as compared with several state-of-the-art techniques.