Self-Similarity Block for Deep Image Denoising

Non-Local Self-Similarity (NLSS) is a widely exploited prior in image denoising algorithms. The first deep Convolutional Neural Networks (CNNs) for image denoising ignored NLSS and were made of a sequence of convolutional layers trained to suppress noise. The first denoising CNNs leveraging NLSS prior were performing non-learnable operations outside the network. Then, pre-defined similarity measures were introduced and finally learnable, but scalar, similarity scores were adopted inside the network. We propose the Self-Similarity Block (SSB), a novel differentiable building block for CNN denoisers to promote the NLSS prior. The SSB is trained in an end-to-end manner within convolutional layers and learns a multivariate similarity score to improve image denoising by combining similar vectors in an activation map. We test SSB on additive white Gaussian noise suppression, and we show it is particularly beneficial when the noise level is high. Remarkably, SSB is mostly effective in image regions presenting repeated patterns, which most benefit from the NLSS prior.