Bent Identity-based CNN for Image Denoising

Abstract

ABSTRACT In the process of image acquisition and transmission, the image will be polluted by noise. Therefore, we propose a bent identity-based convolutional neural network (BICNN) model. The model is a full convolutional network model with a depth of 30 layers, consisting of six feature extraction modules (FEM) and skip connection. Skip connection combines the output features of the first convolution layer with the output features of each FEM in series to guarantee the full extraction of image’s features. Then we adopt the residual learning to alleviate the gradient disappearance and improve the convergence speed so as to ensure that the nonlinear mapping acquired by the trained denoising model is image noise. Bent identity is selected as the activation function, which has soft saturation and the output mean is close to zero, which can enhance the robustness of the model against input noise and accelerate the convergence of the model. Our extensive experiments demonstrate that our BICNN model can not only exhibit high effectiveness in several general image denoising tasks, but also make it highly attractive for practical denoising applications.