摘要： 现有ManTra-Net、DWT-CNN等基于深度学习的数字图像篡改检测算法存在计算复杂度高、检测准确度低等问题。为提取图像篡改与非篡改区域的差异性特征，提出一种基于MobileNetV3-LSTM混合模型的图像篡改检测算法。采用双分支网络架构，主分支网络为带有空洞卷积的轻量级CNN特征提取网络，副分支网络学习篡改图像边界上的差异性，在融合多尺度特征后进行端到端训练，最终输出预测定位图。在COVERAGE、CASIA2和COLUMBIA标准数据集上的实验结果表明，与Xavier-CNN、ELA等算法相比，该算法检测准确度平均提高9.2个百分点，参数量缩减82.3%，推理速度加快2倍，并且具有一定的泛化能力，适用于复制-粘贴、拼接等图像篡改操作的篡改区域检测定位任务。

关键词: 图像篡改检测, 轻量级网络架构, 多尺度融合, 边界差异信息, 被动取证

Abstract: The existing deep learning-based algorithms for digital image tampering detection, such as ManTra-Net and DWT-CNN, suffer from the high computational complexity and low detection accuracy.To capture the discriminative features of tampered areas and normal areas in an image, an algorithm for image tampering detection is proposed based on a mixed MobileNetV3-LSTM model.This algorithm employs a two-branch structure, where the main branch is a lightweight CNN with atrous convolutions designed for feature extraction, and the deputy branch is designed to learn the discriminative features around the boundary of tampered images.After multiscale features are fused, end-to-end training is performed to output the predicted positioning of tampered areas.The proposed algorithm is tested on multiple standard datasets, including COVERAGE, CASIA2 and COLUMBIA.Results show that compared with Xavier-CNN, ELA and other algorithms, the proposed algorithm increases the detection accuracy by an average of 9.2 percentage points, and reduces the number of parameters by 82.3%.This algorithm doubles the reasoning speed, and displays a certain degree of generalization ability, which makes it applicable to the detection and positioning of multiple types of tampered areas, such as those caused by copy-paste and splicing.

Key words: image tampering detection, lightweight network architecture, multiscale fusion, boundary discriminative information, passive forensics

中图分类号: 

• TP391.41