Unpaired Shadow Removal with a Skip-Connected Generator in CycleGAN Framework

This paper proposes an unpaired shadow removal method based on the CycleGAN framework, in which the generator is structurally enhanced by introducing skip connections. Existing CycleGAN-based methods commonly rely on plain encoder-decoder generators originally tailored for style transfer, which often fail to preserve pixel-level spatial alignment and low-level structural details. This limitation leads to incomplete shadow removal, texture degradation, and inconsistent illumination in the restored images. To overcome these issues, the proposed generator introduces skip connections that inject low-level features directly from the encoder to the decoder. This architectural design helps preserve structural integrity and improve visual coherence across the entire image. By leveraging both global and local information, the network achieves a more accurate reconstruction of shadow-free images. The method is evaluated on the public ISTD dataset. Quantitative results show that it consistently outperforms existing unpaired models including CycG, Mask-CycG, and DC-CycG, measured by root mean square error (RMSE), peak signal-to-noise ratio (PSNR), and structural similarity index (SSIM). Visual comparisons further confirm fewer artifacts, more complete shadow removal, and more consistent illumination in the restored regions. Feature map analysis further confirms that the skip-connected architecture captures richer and more diverse semantic features, facilitating better decoding. These findings highlight the practical and theoretical value of enhancing CycleGAN-based generator with skip connections, offering an effective solution for high-quality unpaired shadow removal.