Siamese Differential AE-Net: Noisy Patch Comparison and Its Application in NLM

The structural similarity of image patches is widely used in image processing because of its simple calculation and effectiveness. However, the calculation of image patch similarity is often affected by noise, which makes it difficult to adjust the parameters for image processing tasks, such as the image denoising, classification etc. In this paper, we propose a Siamese differential Auto-Encoder network (AE-net), which applies the deep features of the noise patches to the noise patches comparison. The self-learning characteristic of the AE-net helps to adjust the parameters and extract the deep features of similar patches according to the cost function. The proposed cost function can effectively reduce the difference between the deep features of the noisy and noise-free images. We apply our scheme to the Non Local Means (NLM) denoising method. The experimental results show that our method outperforms the traditional NLM in terms of the PSNR and SSIM values and can better preserve the texture detail of the image while denoising.