摘要
传统的篡改方法如拷贝粘贴和拼接已演变为利用深度学习生成的高质量伪造图像,这些篡改技术在图像纹理和细节上留下难以察觉的痕迹,如高频噪声模式的异常、颜色分布的微妙变化,以及边缘区域的不自然过渡。这些痕迹分布在不同分辨率层次和空间位置,增加了检测的难度。现有模型在整合多尺度和多位置特征时存在不足,难以有效捕捉局部细微纹理变化。针对这一问题,文中提出一种基于多分支HRNet的图像篡改检测与定位模型。该模型通过集成纹理增强模块,增强对图像篡改细节特征的捕获能力。同时,结合Spatial Weighting与Cross Resolution Weighting策略优化特征融合,并使用新的损失函数W_Arcloss,显著提升了模型在复杂篡改检测任务中的性能。在CASIA、Columbia、COVERAGE和NIST16等数据集上,该模型的检测准确度相较于PSCC⁃Net、HIFI⁃Net模型分别平均提升了6.5%与0.8%,并且泛化能力得到提升。这些结果证明了模型在处理多种篡改类型时的有效性和鲁棒性,为图像篡改检测与定位领域提供了新的研究视角和技术手段。
Traditional tampering methods such as the copy and paste and the stitching have evolved into high⁃quality forged images generated by deep learning.These tampering technologies leave imperceptible marks on the textures and details of the image,such as anomalies in high⁃frequency noise patterns,subtle changes in color distribution,and unnatural transitions in edge regions.These traces are distributed at different levels of resolution and spatial positions,which increases the difficulty of detection.The existing models have shortcomings in integrating multi⁃scale and multi⁃position features,which makes it difficult to capture local subtle texture changes effectively.In view of the above,this study proposes an image tamper detection and localization model based on multi⁃branch HRNet(high⁃resolution net).This model enhances the ability to capture the image tampering details by integrating a texture enhancement module.In addition,the feature fusion is optimized by combining the strategies of Spatial Weighting and Cross Resolution Weighting,and a new loss function W_Arcloss is adopted,which significantly improves the model performance in complex tasks of tamper detection.On datasets such as CASIA,Columbia,COVERAGE and NIST16,the detection accuracy of this model has improved by an average of 6.5%and 0.8%in comparison with the PSCC⁃Net and HIFI⁃Net models,respectively,and its generalization ability has been improved.These results demonstrate the effectiveness and robustness of the model in processing multiple types of tampering,which provides a new perspective of research and technical means for the field of image tampering detection and localization.
作者
曾桢
谭平
ZENG Zhen;TAN Ping(School of Information,Guizhou University of Finance and Economics,Guiyang 550025,China;Big Data Institute,Wuhan University,Wuhan 430000,China)
出处
《现代电子技术》
北大核心
2025年第3期35-42,共8页
Modern Electronics Technique
基金
国家自然科学基金项目(71964007)。
