Shadow areas in document images,which are prone to being influenced by lighting,can seriously affect users'recognition and reading.To address the issues of limited shadow removal open datasets and poor quality of corrected image shadow elimination,an improved Generative Adversarial Network(GAN)approach for document image shadow removal is proposed.This network introduces a shadow-aware directional context module and a mask attention structure based on the original dual-GAN architecture.An adaptive attention module is further appended to the mask attention module,which utilizes a parallel structure to adaptively adjust the convolution kernel size and fuse features of different scales to obtain more feature information.Additionally,a pixel-wise superimposition data augmentation method is adopted to solve the scarcity of document shadow removal datasets.Experimental results demonstrate that compared with DCGAN,ST-CGAN,and DSC methods,the proposed approach achieves improvements in evaluation metrics such as Peak Signal-to-Noise Ratio(PSNR)and Structural Similarity Index Measure(SSIM)on the self-constructed dataset.
document image shadow removalsackedshadow direction is context-awaremask attention
维普
万方数据
