Wavelet-Based Loss Function for Image Denoising by Neural Networks

Image processing related to noise suppression may entail partial distortion of useful visual information caused by filtering. Preservation of contours and texture details in the process of image denoising is one of the important problems of computer vision. Recent years deep neural networks are widely applied in the tasks of noise suppression and signal recovering. However, neural networks are subjects to spectral bias towards low-frequency components. It results in blurring of contours and local high-frequency details after filtration. We propose loss function based on using high-frequency information of wavelet representation of images during network training to solve the problem. Our experiments demonstrate that this approach improves results of filtration in terms of signal to noise ratio and structural similarity index.