Adaptive Aggregation of Multiple Denoisers with Mismatched Correlated Noise Models

Abstract

Denoisers for stationary correlated noise require explicit or implicit knowledge of the noise power spectral density (PSD). However, the PSD characterizes the noise correlation only as long as this is stationary. Yet, even when images are corrupted by noise with nonstationary correlation, denoisers can still assume a specific PSD, and successfully filter locally on regions where the noise correlation matches the model provided by the PSD. Elsewhere, the denoising is ineffective and it can introduce distortions and artifacts. One may then want to combine multiple denoised images obtained for a bank of PSDs spanning the diverse correlations found locally over the image, such that each pixel of the final output is obtained using a relevant PSD. However, when the nonstationary correlation is signal-dependent—as in many cases of practical relevance—this combination is not trivial, because the signal is unknown, and so are the regions with matched noise model. We introduce an effective aggregation method specifically designed for signaldependent correlation, where locally adaptive weights reward the statistical compatibility of estimates and corresponding residual noise with the assumed model. Our adaptive aggregated estimate is close visually and in terms of PSNR to that with oracle weights based on the noise-free image.