A Robust Image Denoising Method With Multiview Texture-Aware Convolutional Neural Networks

Image denoising is an issue of intensive research in the image processing community. As the wave of deep learning advances, image denoising with convolutional neural networks has recently made drastic progress, however, they usually can not recover atypical details from noisy images. Motivated by the above, we propose a multiview texture-aware convolutional neural network named MVTANet, which comprises primary denoising network and multiview texture-aware modular. Proposed multiview texture-aware modular owns two variants, i.e., main and secondary texture-aware modular. First, the denoised image is obtained through the primary denoising network. Then, the denoised image and clean image are input into multiview texture-aware modular, respectively, and we obtain two sets of intermediate features and calculate the corresponding perceptual loss. This perceptual loss is designed to generate auxiliary supervision for tiny detail recovery. By setting different initial parameters and parameter freezing technology, this modular can be further focused on restoring atypical details. Extensive experiments demonstrate that the proposed MVTANet is superior to the state-of-the-art denoising methods.

Image denoising is an issue of intensive research in the computer vision and image processing community since it plays a crucial role in applying high-level vision tasks in the subsequent research. Its goal is to recover a clean image from corrupted image with various additive noises. At present, the additive white Gaussian noise (AWG) is one of the most common noise scene, so we assume that n is drawn from a Gaussian distribution with mean 0 and variance $\sigma ^{2}$, i.e., $n \sim N (0,\sigma ^{2})$. Over a period of time, single image denoising methods with AWG noise have been studied by most scholars and have recently made drastic progress.^– For example, nonlocal means (NLM), block-matching 3-D filtering (BM3D), weighted nuclear norm minimization (WNNM), a cascade of shrinkage fields (CSF), trainable nonlinear reaction diffusion (TNRD), and so on are representative image denoising methods. These traditional methods recover the clean image by mainly exploiting the global or nonlocal similar information from the current single noisy image, or using prior knowledge to smooth the noisy image and recover the details of a potentially clean image to a certain extent. However, the existing traditional single image denoising method has the following shortcomings and limitations: 1) most methods need to be further optimized in the test stage; 2) the parameters in the model need to be set manually; 3) not enough training data, leading to poor universality of the model. As the wave of deep learning technology advances, its strong self-learning ability is not only prominent in advanced visual tasks,^, but also applied in various low-level computer vision tasks such as image superresolution, image denoising, and image restoration.^– In 2012, Burger et al. first introduced the multilayers neural network to solve image denoising tasks superior to BM3D, which owned the most advanced performance at that time. Zhang et al. designed DnCNN from another point of view, which focuses on learning noise images rather than directly predicting the repaired image. Lefkimmiatis made full use of the inherent nonlocal self-similarity of natural images, and proposed a variational method with a nonlocal depth network model for image denoising. For the different noise levels, FFDNet can use the same model to solve the blind denoising task. For the real image with unknown noise, CBDNet with a noise estimation network can effectively estimate the noise map in different levels and remove noise from noisy images to achieve blind denoising of the image.