Lightweight parameter de-redundancy demoiréing network with adaptive wavelet distillation

Recently, deep convolutional neural networks (CNNs) have achieved significant advancements in single image demoiréing. However, most of the existing CNN-based demoiréing methods require excessive memory usage and computational cost, which considerably limit to apply on mobile devices. Additionally, most CNN-based methods employ expensive approaches to generate similar feature maps, thereby resulting in redundant parameters within these networks. To alleviate these issues, we propose a lightweight parameter de-redundancy network (PDNet) for image demoiréing. Specifically, we present an efficient ghost block (EGB) that utilizes dilated convolution and cost-efficient operation, which significantly reduces parameters and extracts representative features. Meanwhile, we design a multi-scale shuffle fusion mechanism (MSFM) with a low amount of parameters to integrate different scales of features, which mitigates the information loss issue. To enable the lightweight network for learning latent moiré removal knowledge better, we adopt adaptive wavelet distillation to guide the network training. Experimental results validate the efficacy of our proposed method, achieving comparable or even superior results to the state-of-the-art method, while utilizing only 26.32% of parameters and 29.59% of Macs. The source code will be publicly available at https://github.com/ChenJiaCong-1005/PDNet .