TSPLNet: a three-stage progressive lightweight network for shadow removal

Shadow removal is a significant yet challenging task in computer vision. Most existing methods for shadow removal rely on “black-box” models, which lack transparency and fail to fully integrate shadow removal theory. Moreover, their large number of parameters renders them unsuitable for scenarios with limited computational resources. This study introduces TSPLNet, a lightweight and progressive network for shadow removal, combining a physics-based model with a data-driven approach. First, we incorporate shadow removal principles into image restoration theory to redefine the shadow removal process. Based on this new computational framework, we design a three-stage progressive deep learning network, where each stage iteratively refines shadow reconstruction. To improve the model’s ability to extract and reconstruct shadow features, we introduce key components such as a deformable residual block, a shadow attention adjuster, and a shadow interaction attention module, along with a redesigned loss function. Furthermore, we apply depthwise separable convolution, reducing the model’s parameters to 36% of the original. Experiments on the ISTD and SRD datasets, as well as crop leaf images demonstrate that our method significantly outperforms other advanced methods in terms of parameter count, quantitative metrics, and visual quality, highlighting its effectiveness in image shadow removal.