Abstract: Image tampering detection has become critically important for digital forensics and media content verification. However, in real-world applications, images often undergo post-processing such as brightness and contrast adjustments, which can weaken tampering traces and degrade the performance of existing algorithms. To address this challenge, a novel image tampering detection network ReConWave-Net is proposed. The network consists of two key modules: a classification-guided restoration module that adapts to specific post processing types to counteract brightness and contrast distortions, and a tampering localization module that leverages multi-scale wavelet features and a contrastive learning mechanism to enhance the representation and localization of forged regions. The method was evaluated on multiple datasets under various brightness and contrast perturbations. In terms of restoration quality, the average PSNR increased from about 10 dB to 31 dB, and SSIM improved from 0.40 to 0.92; under typical perturbations, the detection performance reached an F1 score of 0.73 and an IOU score of 0.65. The results indicate that combining targeted restoration with detection can significantly enhance the robustness of forgery localization in post-processed images.

Key words: image forensics, imaestoration, tampering detection, contrastive learning, wavelet feature enhancement

摘要： 数字图像篡改检测在数字取证和媒体内容验证等领域具有重要意义。然而，实际应用中篡改图像经常经历亮度、对比度等后处理操作，这会削弱篡改痕迹并削弱现有算法的检测性能。针对这一问题，本文提出了一种新颖的图像篡改检测网络——ReConWave-Net。该网络包含2个关键模块：分类引导的图像恢复模块用于根据图像扰动类别对图像进行针对性恢复，以减弱亮度和对比度扰动的影响；篡改定位模块通过多尺度小波特征和对比学习机制，强化了篡改区域的特征表达和定位能力。在多数据集及若干亮度、对比度扰动下评估方法，恢复质量方面平均 PSNR 由约 10 dB 提升至31 dB，SSIM 由0.40 提升至0.92；检测性能在典型扰动下F1分数0.73、IOU分数0.65。结果表明，将针对性恢复与检测相结合，可显著提升对后处理图像的篡改定位鲁棒性。

关键词: 图像取证, 图像恢复, 篡改检测, 对比学习, 小波特征增强