.png){kind=icon}
.png){kind=icon}
Return
Transformer-enhanced U-Net generator for unpaired shadow removal
.png){kind=icon}
PRE
AI
L
C
K
DOI：10.1117/1.JEI.35.2.023023.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Faithful shadow removal in unpaired settings remains challenging because the absence of pixel-level supervision often leads to artifacts or structural inconsistencies. To compensate for this limitation, many unpaired approaches introduce complex architectures or auxiliary modules, which improve performance but substantially increase computational cost. Meanwhile, the convolutional neural network-based generators commonly employed in unpaired frameworks operate under inherently local receptive fields, limiting their ability to capture long-range dependencies and global illumination relationships. We present a lightweight transformer-enhanced generator within a CycleGAN framework for unpaired shadow removal. The generator adopts a compact skip-connected U-Net backbone and integrates a Local-Global Feature Module at each stage, where lightweight local refinement is achieved through a single-layer convolution and 1D channel attention. A Swin Transformer bottleneck is further incorporated to strengthen long-range dependency modeling and promote global illumination consistency. This design jointly enhances local fidelity and global semantic coherence while keeping the overall model compact and efficient. The framework remains fully unpaired and requires neither paired ground truth nor manually annotated masks. Extensive experiments on ISTD and cross-domain evaluation on SRD demonstrate that the proposed method achieves competitive results against representative baselines, while using over 160 & times; fewer parameters.
Keywords:
shadow removal
unpaired learning
Cycle GAN
Swin transformer
shadow masks
Journal
J
IF:
1
Papers:
109
Citations:
2.7K
Organization
Cited Papers