An adaptive CNN for image denoising

Convolutional neural network techniques have shown great promise in many low-level image processing tasks. However, due to the reduced impact of convolution operations in deep layers on the original inputs, the image denoising task suffers from effective information loss during forward propagation. To address this issue, this paper proposes an adaptive denoising CNN (ACNN) with a deep feature block (DFB), an adaptive block (AB), and a construction block (CB). The DFB utilizes a stacked architecture to ensure the learning ability of ACNN. The AB dynamically adjusts its learning strategy by using a deformable convolutional filter and a switchable normalization (SN) to extract more powerful features. CB gathers DFB and AB to extract complementary features for images. Also, CB can be used to construct a clean image. Moreover, a consolidate loss function utilizes mean squared error (MSE) and structure similarity index measure (SSIM) to further alleviate smooth problem. Experimental analysis shows the superiority of the proposed ACNN in terms of qualitative analysis and quantitative metrics.