A lightweight CNN based information fusion for image denoising

Deep convolutional neural networks (CNNs) with strong learning abilities have obtained good results for image denoising. However, the CNNs for image denoising have increasingly heavy-weighted, which is not suitable for practical applications. In this paper, we present a lightweight width information fusion CNN(LWIFCNN) for image denoising to address this problem. The proposed model employs two key modules, i.e., multi-scale width information block (MWIB) and information enhancement block (IEB), to improve the model representing capability without heavy complexity. Specifically, in order to extract as much information as possible with low weights, MWIB utilizes a standard convolution and three lightweight residual attention blocks (RABs) to achieve multi-scale feature fusion. Each RAB utilizes two lightweight blocks (LWBs) and an enhanced channel attention mechanism (ECA) to extract width information and reduce computational complexity. IEB uses serial modules and ghost modules to combine width features and depth features to further enhance the representing capability of the model. Experimental results show that our method is better than many excellent methods in both quantitative and qualitative metrics.