基于方向感知与注意力的文档图像阴影消除

Research on Shadow Removal of Document Image based on direction sensing and Attention Mechanism

文档图像易受光照影响产生的阴影区域,会严重影响用户的识别和阅读。针对现有阴影消除开放数据集少,纠正图像阴影消除质量较差的问题,提出了一种改进的生成对抗网络的文档图像阴影消除方法。该网络在原有双GAN网络结构基础上,引入阴影感知方向上下文模块以及掩码注意力结构,并在掩码注意力模块后加入自适应注意力模块,通过并行结构自适应调整卷积核大小,融合不同尺度特征以获得更多特征信息。同时采用一种逐像素叠加的数据增强方法,解决文档阴影消除数据集缺乏的问题。实验结果表明在本文自建数据集上与DCGAN、ST-CGAN和DSC方法相比,评价指标PSNR与SSIM均有一定提升,本文实验方法优于现有文档去阴影先进方法。

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.