多向堆叠记忆网络在证件图像篡改检测中的应用  

Application of Multi-direction Stacked Memory Network in Certificate Image Tamper Detection

作　　者：赵卫东[1] 黄见 张睿[1] 吴乾奕 ZHAO Weidong;HUANG Jian;ZHANG Rui;WU Qianyi(School of Software,Shanghai Key Laboratory of Data Science,Fudan University,Shanghai 200433,China)

机构地区：[1]复旦大学软件学院上海市数据科学重点实验室,上海200433

出　　处：《小型微型计算机系统》2025年第2期346-352,共7页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(71971066)资助。

摘　　要：随着金融线上业务的迅猛发展,篡改图像信息的问题在风控环节频繁出现.然而,现有的篡改检测模型在处理证件图片的准确性和应对环境干扰方面亟需加强.为解决这一问题,本文提出了一种二阶段篡改检测模型:在第1阶段中,通过将简单堆叠长短期记忆网络改进为多方向堆叠记忆网络,弥补了篡改特征对比方向单一的问题,并且兼顾了图像的位置信息,从而提高篡改鉴别准确率.第2阶段是在初步确定篡改区域后,基于篡改区域外围多层邻域的纹理特征,以注意力机制为核心推测中心区域纹理特征值,再与原中心区域纹理特征值对比筛选假阳性区域.实验表明,本文的改进方法是有效的.With the rapid development of online financial bussinesses,the issue of tampering with image information frequently arises in the risk control process.However,existing tamper detection models urgently need improvement in the accuracy of handling certificate images and addressing environmental interference.To address this problem,this paper proposes a two-stage tamper detection model,TSTD.In the first stage,a simple stacked long short-term memory network is improved to a multi-directional stacked memory network.This improvement addresses the single problem of tamper feature contrast direction,considers the location information of the image,and enhances the accuracy of tamper detection.The second stage involves inferring the texture features of the central area based on the texture features of the multi-layer neighborhood around the tampered area.Subsequently,the false-positive areas are filtered by comparing the texture features of the original central area.Experimental results demonstrate the effectiveness of these methods.

关 键 词：篡改检测 证件图像 多向堆叠记忆网络 多邻域纹理特征 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]