摘要
现有文本图像阴影消除方法已取得了一定的进步,但是这些方法主要关注图像本身和文本背景颜色信息,忽视了真实场景中文本图像通常伴随密度不一致的阴影,因此这些方法可能存在图像局部细节丢失、内容不协调等问题.针对这个问题,本文提出了一种基于密度和颜色信息指导的文本图像阴影消除算法.首先,设计了特征提取模块,以提取输入图像的阴影密度和背景颜色信息.随后,在每一级模块输入之前,利用自适应模块调整初始的阴影密度特征,并结合背景颜色信息指导后续的阴影消除过程.为了更好地提取图像的全局与局部特征,本文提出了密度和颜色引导的Transformer模块和阴影敏感的局部特征提取模块,并将两者结合作为主要阴影消除模块.实验结果表明,相比现阶段的文本图像阴影消除方法,所提出的网络模型在性能上更加优越.
Existing methods for shadow removal in document images have made certain progress.However,these approaches primarily focus on the image itself and the background color information,neglecting the fact that in real-world scenarios,document images often contain shadows with varying densities.Consequently,these methods may result in issues such as loss of local details and content inconsistencies.To address these challenges,a document image shadow removal algorithm guided by density and color information is proposed.A feature extraction module is first designed to extract shadow density and background color information from the input image.Before each module receives its input,an adaptive module adjusts the initial shadow density features,which are then combined with the background color information to guide the subsequent shadow removal process.To better extract both global and local features,a densityand color-guided Transformer module,along with a shadow-sensitive local feature extraction module,is proposed.The two modules are combined to form the primary shadow removal module.Experimental results demonstrate that,compared to current methods,the proposed network model outperforms existing text image shadow removal techniques.
作者
柏畅
张玲
BAI Chang;ZHANG Ling(School of Computer Science and Technology,Wuhan University of Science and Technology,Wuhan 430065,China)
出处
《计算机系统应用》
2025年第7期152-162,共11页
Computer Systems & Applications
基金
湖北省自然科学基金(2023AFB615)。
关键词
文本图像阴影消除
深度学习
阴影密度指导
背景颜色指导
自适应调整
document image shadow removal
deep learning
shadow density guidance
background color guidance
adaptive adjustment
作者简介
通信作者:张玲,E-mail:zhling@wust.edu.cn。
