Analysis of the residual with autocovariance

The evaluation method proposed here follows the idea presented in [Riot et al., 2017] which examines the residual image and looks for possible remaining structural elements in this residual image. Instead of using autocorrelation, we used patch base autocovariance method during the residual evaluation. Unlike maximum ENL estimation or αβ estimation [Gomez et al., 2016] method, this method is automatic and does not rely on a supervised selection of homogeneous regions. It also provides a global score for the whole image.

Figure 1. Denoising real Sentinel-1 images over the region of Saclay (the original noisy image is available in figure 6.2(a)). Left column : denoised results ; middle column : residual ratio images ; right column : residuals evaluation results with displaying value range [0, 4]. 64 Sentinel-1 images are used.

This method can highlight the areas, which still contain some textures, in the ratio image between the denoised data and the noisy data.

Code links

Residual-evaluator: under preparation