文字图像不规则干扰修复算法研究  

Irregular Interference Inpainting Algorithm Research on Text Image

作　　者：段荧 龙华[1,2] 瞿于荃 杜庆治[1,2] 邵玉斌[1,2] DUAN Ying;LONG Hua;QU Yu-quan;DU Qing-zhi;SHAO Yu-bin(College of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650000,China;National Key Laboratory of Computer Science of Yunnan Province,Kunming University of Science and Technology,Kunming 650000,China)

机构地区：[1]昆明理工大学信息工程与自动化学院,昆明650000 [2]昆明理工大学云南省计算机重点实验室,昆明650000

出　　处：《小型微型计算机系统》2021年第7期1427-1434,共8页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61761025)资助。

摘　　要：该文针对不规则干扰导致文字图片字符识别率下降的问题,提出一种基于U型网络框架和部分卷积运算的文字图片修复模型.首先,针对常见字体的干扰问题,通过图像融合建立干扰文字图像数据库,在逐像素损失、感知损失和全变分损失的共同约束下,根据已有笔画细节对污损部分进行修复,并对污损汉字的字体形状和笔画走向的细部特征进行复原;其次,使用光学字符识别接口对修复前后图片进行测试并计算识别率;最后,将该文算法初步应用于真实场景下的古代文字拓片修复.实验证明,该文模型在常见文字修复上峰值信噪比最高达到32.58 d B,最佳损失值为0.015,污损文字图片修复后识别准确率提升30.49%.To solve the character recognition-rate declining problem with text pictures caused by irregular interference,a inpainting model is proposed here for rebuilding a text picture based on partial convolution layers and U-Vet framework.Firstly aiming at the interference problem of some common fonts,a database of interfering text images is established through image fusion.Under the combined constraints of pixel-by-pixel loss,perceptual loss,and full-variable fraction loss,repair the defaced parts based on the existing stroke details,and recover the fine features of the font shape and stroke direction of the defaced Chinese characters.Secondly through the optical character recognition API,the degrees of accuracy change in the picture before and after restoration are calculated.Finally,a preliminary application of the text algorithm to ancient text topography restoration in real-life scenarios.As proved by tests,the proposed model had a peak signal-to-noise ratio of 32.58 dB and the best loss value was 0.0015 on common text restoration,the recognition accuracy increased by 30.49%after the inpainting of the defiled and defaced text pictures.

关 键 词：文字图像修复 部分卷积 光学字符识别 深度学习 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]