LMENet: A lightweight multiscale efficient convolutional neural network for image denoising

Various convolutional neural network (CNN) based image denoising methods have been proposed and achieved their attractive denoising performances in past decades. These methods benefit from carefully designed network structures via their increasing network depth and width. Since enlarged networks suffer from numerous parameters and cause heavy computational burden, their deployment application is limited on portable devices. To tackle this problem, we propose a lightweight multiscale efficient convolutional neural network (LMENet). It uses a set of cascaded residual networks (RNs) as its backbone network, where multiscale features are exported from these RNs. By fusing these features with a dual attention mechanism, denoised images are finally reconstructed. The experiments show that, compared with some state-of-the-art methods, our LMENet has a competitive performance for denoising in both additive white Gaussian noise (AWGN) and real noise tests. More importantly, it has approximate one-eighth of the parameter size compared with the existing methods. It demonstrates the computational efficiency of the LMENet, while achieving similar PSNR results.