摘要： 基于深度学习的图像复制-粘贴篡改检测方法在特征提取过程中未考虑特征的空间排列，在小区域篡改样本下检测性能不佳。基于可形变自相关网络提出一种图像篡改检测方法。通过引入可形变卷积和多尺度空间金字塔，自适应地学习篡改目标的空间形变，同时通过构造自相关金字塔式特征层次结构，融合全局特征和局部特征以提升图像篡改检测性能。实验结果表明，该方法在公开的图像篡改检测基准上各项评测指标均优于对比方法，其精确率、召回率、F₁值较BusterNet 2019分别提高14.85、15.04、12.81个百分点，在小区域篡改样本下性能优势更为明显。

关键词: 图像篡改检测, 特征提取, 可形变卷积, 自相关金字塔

Abstract: The deep learning-based copy-move forgery detection methods ignore the spatial layout of the features, leading to a reduction in the detection performance for small-region forgery samples.Additionally,the fixed size of the receptive fields in Convolutional Neural Network (CNN) modules are not suitable for the detection of nonrigid image forgery.To address the problem,a copy-move forgery detection method is proposed based on a deformable self-correlation network.The method introduces deformable convolution and multi-scale spatial pyramid to adaptively learn the spatial deformation of the forgery target.At the same time,a self-correlation pyramidal feature hierarchy is constructed to integrate the global features and local features to improve the performance of image forgery detection. Experimental results show that the proposed method is superior to the compared methods in all indexes of image forgery detection performanc.Compared with BusterNet 2019,the proposed method increases the accuracy by 14.85 percentage point,recall rate by 15.04 percentage point,and F₁ score by 12.81 percentage point, especially in the case of small-region forgery samples.

Key words: image forgery detection, feature extraction, deformable convolution, self-correlation pyramid

中图分类号: 

• TP391