Lab-DN: Dual-branch lightweight network for Lab color space shadow removal

This paper aims to address the limitations of current over-parameterized shadow removal models and proposes a novel dual-branch lightweight deep neural network for processing shadow images in the Lab color space. The proposed network, called “Lab-DN”, is motivated by three key observations. Firstly, the Lab color space effectively separates luminance information and color properties, which have not been fully utilized by existing networks. Secondly, the sequential stacking of convolutional layers fails to leverage features from different receptive fields. Finally, non-shadow regions contain important prior knowledge for mitigating the significant color differences between shadow and non-shadow regions. Therefore, we design our Lab-DN with a dual-branch structure, consisting of “L” and “ab” branches, where shadow-related luminance information is processed in the L branch, while the ab branch retains chrominance properties. Moreover, each branch comprises several Multi-receptive-fields Information Fusion modules (MIF), Shadow Boundary-aware Attention modules, and convolutional filters. Each MIF module includes multiple parallelized dilated convolutions with varying dilation rates to receive different receptive fields and operate with distinct network widths, thereby reducing model computational costs. In addition, a Laplacian-filter-based Channel Attention module is incorporated to aggregate features from different receptive fields and achieve better shadow removal performance. Extensive experiments on the ISTD, ISTD+, and SRD datasets demonstrate that our Lab-DN achieves +0.22/+0.18/+0.76 PSNR improvements-in ISTD improving from 36.95 dB to 37.17 dB compared to EMDN. Our model has only 0.93 M parameters and requires approximately one-tenth the computational cost.