True wide convolutional neural network for image denoising

Abstract

Deep convolutional neural networks (CNNs) have achieved significant success in image-denoising tasks. However, increasing the network depth arbitrarily results in vanishing/exploding gradients, which does not improve the image denoising quality. In this study, we propose a true wide CNN (WCNN) to reorganize several convolutional layers. Through wavelet decomposition, all convolutional layers that are originally concentrated on a single information stream to train one image are now distributed among multiple independent subnetworks, each of which has few convolutional layers and is arranged in parallel in the WCNN. Each subnetwork has its own input and output and is supervised by its own loss function to capture image features with a specific direction and scale, allowing the WCNN to have sufficient convolutional layers to capture image features while avoiding vanishing/exploding gradients. In addition, each subnetwork can adopt a suitable CNN structure based on the characteristics of a subband, ensuring that the WCNN is competent in handling a specific range of noise levels using a single set of trained parameters. Extensive quantitative and qualitative evaluations of benchmark datasets and real images reveal that the WCNN outperforms state-of-the-art denoising methods.