A lightweight and effective neuron attention convolutional neural network for image denoising
Jibin Deng · 2023
Noise and image details in a noisy image are highly correlated, making image denoising as a task that involves noise removal and image details retention. However, their correlation is rarely considered, causing unsatisfactory denoising performance. To this end, a lightweight and effective neuron attention denoising network (NADNet) is proposed, whose denoising effectiveness is attributed to mixed dilated convolutions block (MDCsB), neuron attention block (NAB) and residual learning (RL), in which MDCsB helps NADNet capture more noise and alleviate artifacts, by stacking different dilated convolutions with inconsistent dilated factors, NAB facilitates NADNet to focus more on valuable image features for effectively capturing noise and retaining more original image details, and RL helps NADNet better tackle deeper neural network training difficulties and reduce over-fitting. Experimental results show that NADNet achieves superior denoising performance.