A novel all-reference image denoising quality evaluation method based on joint sparse representation

Numerous digital image algorithms have been proposed along with the emergence of new signal representation methods. However, denoising quality evaluation attracts less attention yet. In this paper, a novel all-reference image denoising quality evaluation method based on Joint Sparse Representation (JSR) is proposed. In this method, we model the clean image and the corresponding denoising image as a group images sensing the same scene but have differentiated contents. By joint sparse decomposition, these images are separated as a common part and two differentiated private parts; each of them is a linear combination of atoms in the given learning dictionary and stand for different aspects of image content. With this distinguishable information, we can distinguish image detail loss and noise remnants, and then evaluate the denoising quality from these two aspects. Compared with traditional statistical evaluation index, our method provides a more informative measure method in sparse feature space. Extensive experimental results demonstrate the superiority of our method.