Improved Denoising Auto-encoders for Image Denoising

Image denoising is an important pre-processing step in image analysis. Various denoising algorithms, such as BM3D, PCD and K-SVD, obtain remarkable effects. Recently a deep denoising auto-encoder has been proposed and shown excellent performance compared to conventional image denoising algorithms. In this paper, we study the statistical features of restored image residuals produced by Denoising Auto-encoders and propose an improved training loss function for Denoising Auto-encoders based on Method noise and entropy maximization principle, with residual statistics as constraint conditions. We compare it with conventional denoising algorithms including original Denoising Auto-encoders, BM3D, total variation (TV) minimization, and non-local mean (NLM) algorithms. Experiments indicate that the Improved Denoising Autoencoders introduce less non-existent artifacts and are more robustness than other state-of-the-art denoising methods in both PSNR and SSIM indexes, especially under low SNR.