【作者】 张鹏飞；

【导师】 李永；

【作者基本信息】 北京邮电大学 ， 电子科学与技术， 2025， 硕士

【摘要】 印章图像具有重要的法律效力,伪造印章图像可能引发重大的经济和社会风险。然而,在印章图像的真伪鉴别过程中,常面临三大挑战:1)背景干扰(印章本体与文字或签名重叠)导致印章特征提取偏差;2)伪造印章样本数据稀缺;3)伪造印章图像与真实印章图像的视觉差异细微。为此,本文提出了两阶段鉴别框架。在第一阶段,即印章精确提取与几何对齐阶段,基于YOLOv7-tiny实现印章定位,结合自适应双重阈值融合法和SIFT-RANSAC消除背景干扰并完成几何对齐。在第二阶段,即印章图像相似度测度设计与印章真伪鉴别阶段,通过高仿伪造印章数据生成与边缘特征优化,构建SVM分类器实现印章真伪鉴别,从而解决数据稀缺与差异判别问题。本文提出了一种印章精确提取和几何对齐方法。在提取阶段,针对印章图像复杂背景干扰的问题,本文使用YOLOv7-tiny模型自动定位待测试图像中的印章区域,根据边界框坐标将印章区域从待检测图像中裁剪出来,减少与印章无关的干扰信息;并提出了一种自适应双重阈值融合法,通过自适应HSV阈值和通用HSV阈值的融合加权,动态确定印章本体的颜色分布,将非印章本体的像素设置为印章图像背景主导颜色,再利用最大类间方差法将印章本体与背景进行分离得到印章二值图像,消除背景干扰,使得后续真伪鉴别专注于印章本体的特征。在几何对齐阶段,本文采用SIFT和RANSAC算法对印迹二值图像和印模二值图像进行几何对齐,用于消除几何视差,并提出最小误差对齐方案,提升匹配可靠性,便于后续通过差分图像比对印迹二值图像与印模二值图像之间的细微差异。在真伪鉴别阶段,本文建立了一个印章图像真伪数据集,用于后续的二分类模型的训练和评估。其中,在建立数据集过程中,针对伪造印章图像数据稀缺的问题,本文提出了一种高仿伪造印章生成方法,通过OCR与图像处理技术,对印模图像中的文字和数字区域进行缩放、位移和变形等操作,设计了 16种类型的伪造印章生成方法,模拟印章伪造过程可能出现的图像变换。针对真伪印章图像差异细微以及背景元素可能影响真伪鉴别准确性的问题,本文基于印迹二值图像和印模二值图像的差分图像,结合距离变换和细化算法提取印章二值图像的加权矩阵和骨架,改进了基于边缘特征的相似度度量指标(F1至F5),并将这些指标结果作为输入向量,训练了一个SVM二分类模型以鉴别印章图像真伪。实验结果表明,基于F1至F5特征组合的模型能够有效区分高仿伪造印章图像和真实印章图像,其AUC达到了0.995,准确率为96.83%,真阳性率(TPR)为96.30%,假阳性率(FPR)为 2.65%。更多还原

【Abstract】 The seal image carries significant legal validity,and forging seal images may lead to serious economic and social risks.In the process of authenticating the genuineness of seal images,three major challenges are commonly encountered:1)background interference(overlap between the seal body and text or signatures)causing deviations in seal feature extraction;2)scarcity of forged seal sample data;and 3)subtle visual differences between forged and genuine seal images.To address these challenges,this thesis proposes a two-stage identification framework.In the first stage,namely the precise seal extraction and geometric alignment stage,YOLOv7-tiny is employed to locate the seal,and background interference is eliminated and geometric alignment is completed by combining the adaptive dual-threshold fusion method and the SIFT-RANSAC algorithm.In the second stage,namely the seal image similarity measurement design and seal authenticity identification stage,high-fidelity forged seal data generation and edge feature optimization are carried out to construct a SVM classifier for seal authenticity identification,thereby addressing the problems of data scarcity and difference identification.This thesis proposes a method for precise seal extraction and geometric alignment.In the extraction stage,to address the issue of complex background interference in seal images,the YOLOv7-tiny model is employed to automatically locate the seal region in the test image.The seal region is then cropped from the image based on the bounding box coordinates,reducing irrelevant interference information.An adaptive dual-threshold fusion method is also proposed.By combining the adaptive HSV threshold with a general HSV threshold through weighted fusion,the color distribution of the seal body is dynamically determined.Pixels not belonging to the seal body are set to the dominant background color of the seal image.The seal body is then separated from the background using the maximum inter-class variance method to obtain a binary seal image,eliminating background interference and allowing subsequent authenticity identification to focus on the features of the seal body.In the geometric alignment stage,the SIFT and RANSAC algorithms are used to perform geometric alignment between the binary impression image and the binary seal template image to eliminate geometric parallax.A minimum error alignment scheme is also proposed to enhance matching reliability,facilitating the subsequent comparison of subtle differences between the binary impression image and the binary seal template image through differential imaging.In the authentication stage,to address the scarcity of forged seal image data,a high-fidelity forged seal generation method is proposed.Using OCR and image processing techniques,scaling,shifting,and deforming operations are performed on the text and numeric regions in the seal template image.Sixteen types of forged seal generation methods are designed to simulate the image transformations that may occur in the seal forgery process.Subsequently,the generated high-fidelity forged seal images are combined with real seal impression and template images to establish a seal image authenticity dataset for training and validation of the binary classification model.To address the subtle differences between genuine and forged seal images and the potential impact of background elements on the accuracy of authenticity identification,a weighted matrix and skeleton of the binary seal image are extracted from the differential image of the seal impression and template binary images,using distance transformation and thinning algorithms.The similarity measurement indicators(F1 to F5).based on edge features are improved,and the results of these indicators are used as input vectors to train an SVM binary classification model for seal image authenticity identification.Experimental results demonstrate that the model based on the F1 to F5 feature combination can effectively distinguish between high-fidelity forged seal images and genuine seal images,achieving an AUC of 0.995,an accuracy of 96.83%,a true positive rate(TPR)of 96.30%,and a false positive rate(FPR)of 2.65%.更多还原