Shadow Removal Via Global Residual Free Unet and Shadow Generation
Dong Li,Xin Lu,Yurui Zhu,Xi Wang,Jie Xiao,Yunpeng Zhang,Xueyang Fu,Zheng-Jun Zha
DOI: https://doi.org/10.1109/cvprw63382.2024.00634
2024-01-01
Computer Vision and Pattern Recognition
Abstract:Existing shadow removal methods face challenges when confronted with real-world scenes, particularly when dealing with complex background information and high-resolution images. To address these issues, we present a shadow removal framework based on shadow generation and Global Residual Free Unet (GRFUnet), which improves shadow removal from both data and network perspectives. For data enhancement, we train a generative adversarial network for producing shadow masks, which are then multiplied with clean images to obtain new shadow images. The additional shadow data produced by this structured generation approach allows for ample constraining of the network in color aspects, thereby enhancing color consistency in the images. In terms of network architecture, we design the Global Residual Free Unet that employs a convolutional framework to sequentially conduct Spatial Interaction and Channel Evolution for effective feature extraction. Moreover, we eliminate the commonly used Global residual connections in image restoration, as we discern their in-effectiveness for the non-additive task of shadow removal. Through this methodology, we achieve excellent shadow removal results, both qualitatively and quantitatively. Our method achieves the highest PSNR in the NTIRE24-Image Shadow Removal Challenge and achieves commendable and balanced outcomes across two tracks—placing third in the fidelity track and fourth in the perceptual track.