一种古籍文字图像篡改检测识别模型

打开文本图片集

中图分类号：TP751文献标志码：A

A Model for Detection and Recognition of Tampered Ancient Text Images

LI Yongbo 1 ， QIAN Yonggang 2 ， LIU Qin 1 ， MA Yuqi 1 ， WU Sheng 1 ， YU Xianping 1 ， CHEN Shanxiong ⋅1，3 （1.Collge of Computer and Information Science，Southwest University，Chongqing 40O715，China；2. Information Center， ChongqigVocational CollgeofIntellgntEngineeing，Chongqing 40216O，China；3.KeyLaboratoryofEthnic Language Intellgent Analysisand SecurityGovernance，MinistryofEducation，Minzu UniversityofChina，Beijing1Ooo81，China）

Abstract：Toeffectively detectandrecognize tampered textinancientdocument images，atampering detectionand recognition model named MDAS-Net，which canbe used for the character images of ancient texts，was proposed.A fuse atention block was introduced inthe edge-supervised branch to enhance multi-scale feature extraction of imagecontent. Additionally，to improve feature integration between theedge-supervised branch and the noise-sensitive branch，acrossbranch feature transfer modulenamedE-2-N/N-2-EHelp Block wasdesigned，whichfacilitatedeffectiveinformation exchangeand yields higher-qualityfused features.To verifytheefectivenessofthemodel，adatasetofancient textimage tampering was created，and comparative experimentsandablation experimentswereconducted in combination with the Text in Tampered Images （TTI）dataset.The experimental results show that MDAS-Net achieves promising performance in tampered region detection，with an area under curve of receiver operating characteristic（AUC）of O.852 and an F1 （204 score of O.784，confirming its practical value in ancient text image tampering detection.

Keywords： image processing；feature fusion；detection of tampered image；ancient text image；deep learning

在文字图像篡改检测和识别任务中，模型须要通过像素级别的精确定位来区分篡改图像和真实图像，这意味着模型不仅要识别被篡改的区域，而且要精确地定位这些区域。（剩余14253字）