Year: 2024
Type: conference-paper
Abstract: Shadows are a common phenomenon in life, and the removal of shadows from images is a challenging task. Most of the methods for removing shadows use supervised methods, which require paired shadow images and non-shadow images, but obtaining paired datasets is a complex task. This problem is recently addressed by Mask-ShadowGAN proposed by Hu et al. Mask-ShadowGAN is based on Generative Adversarial Network, where the generator and the discriminator fight against each other and constrain each other to achieve shadow removal. However, this network removes shadows with the problems of unsmooth edges and difficulty in recovering information obscured by shadows in complex backgrounds. On this basis, this study integrates the CBAM attention mechanism into the generator network, by adding the CBAM module, the network can prioritize the essential channel features and spatial characteristics of shadow images, so as to improve the performance of the model and the accuracy of shadow removal. Experiments show that on the unpaired dataset USR, compared with other algorithms, the algorithm model proposed in this paper can better retain the original information of the image and have better visual effects. Tested on the paired public dataset ISTD, the algorithmic model proposed in this paper achieves RMSE of 6.94 and PSNR of 27.37 in the overall region of the image. In summary, the algorithmic model proposed in this paper is better at generating shadow-free images close to real environments when compared with mainstream algorithms.
Institution: Jiangxi University of Science and Technology
Language: English
Topic: Image Enhancement Techniques
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4409095035