Learnable nonlocal self-similarity of deep features for image denoising (Q6541916)

The first six pages of the paper provide an excellent literature review on many pros and cons for image restoration when one has all types of noise in the image. This literature review includes the Parzen-Rosenblatt window technique, a statistical approach for calculating the empirical distribution of the given data. The block nonlocal regularizer also is suggested for image patches. In addition, the authors consider the method for nonlocal self-similarity in high-dimensional feature space. They then propose [that] the variational method replaces the block operator with a learnable convolution operator, which greatly reduces computational complexity. The authors also propose a regularizer that can be learned using deep learning methods. The authors then want to capture image features by first transforming the original image into high-dimensional feature space using a feature extraction operator named the head network. They use a popular convolution operator in deep learning and carefully include many details to perform the mission. Numerous additional details are clearly included in the mathematical proofs in several appendices in the paper. The authors then include a numerical scheme for building network architecture by including an algorithm for image denoising by deep feature estimation. They continue by developing the learnable nonlocal self-similarity network for image denoising. By applying the results to six images illustrated in Figures 5 to 10, they illustrate the architecture of the proposed LNSNet images for denoising. They include comparisons of parameters and several routines. The paper includes five appendices containing clearly developed mathematical proofs for the platforms used to show the results. The bibliography contains 50 recent papers developing many of the newer results in the area of deep features for image denoising.