| 摘要: |
| 近年来,随着图像编辑技术的不断发展,数字图像的知识产权保护形势日益严峻。为有效遏制图像内容的非法窃取、复制与篡改行为,切实保障用户的合法权益,图像拷贝检测作为内容审核体系的关键组成部分,其重要性不言而喻。其中,一种常见的规避检测手段是攻击者将源图像的部分内容巧妙地植入目标图像中,形成“画中画”式的局部拷贝。针对这一挑战,本文提出了一种基于ViT的局部图像拷贝检测算法。该算法旨在利用ViT强大的局部特征提取能力,在原始全局特征的基础上提取出图像的局部特征,揭示图像间潜在的依赖关系,并通过重新排列相关图像的顺序,提高具有更高相似度的图像在检测结果中的排序,从而精准实现对图像局部信息的拷贝检测。本文与基于自监督描述符的图像拷贝检测算法(A Self-Supervised Descriptorfor Image Copy Detection)相比,在DISC2021数据集上取得显著提升,局部拷贝图像识别准确率提高10%,μAP (Micro AveragePrecision)指标提高10%～15%。此外,本文还利用热力图进行了直观的定性分析,证实了ViT能够敏锐地捕捉到图像中存在局部拷贝的区域,进一步验证了算法的有效性。本文提出的基于ViT的局部图像拷贝检测重排序算法能对使用画中画图像增强方式的拷贝图像进行有效检测,并在DISC2021数据集上取得了优异成绩,拓展了数字图像取证领域研究的新思路。 |
| 关键词: 图像拷贝检测 ViT 图像增强 重排序 热力图 |
| DOI：10.19363/J.cnki.cn10-1380/tn.2025.11.07 |
| 投稿时间：2024-01-15修订日期：2024-04-19 |
| 基金项目:本课题得到国家重点研发计划项目(No.2020AAA0140000)和浙江省教育厅一般科研项目(No.Y202351231)资助。 |
|
| A ViT-based Algorithm for Localized Image Copy Detection |
| ZHU Chen,CHEN Yuxun,CHEN Yukun,WANG Zonghui |
| Polytechnic Institute, Zhejiang University, Hangzhou 310015, China;College of Computer Science and Technology, Zhejiang University, Hangzhou 310027, China |
| Abstract: |
| In recent years, with the continuous development of image editing technology, the situation of intellectual property protection of digital images has become increasingly severe. In order to effectively curb the illegal stealing, copying and tampering of image content, and effectively protect the legitimate rights and interests of users, image copy detection, as a key component of the content audit system, is of great importance. One of the common means to avoid detection is that the attacker skillfully implants part of the source image into the target image, forming a "picture-in-picture" type of partial copy. To address this challenge, this paper proposes A ViT-based Algorithm for Localized Image Copy Detection. The algorithm aims to utilize the powerful local feature extraction capability of ViT to extract the local features of an image on the basis of the original global features, to reveal the potential dependencies between the images, and to improve the ordering of images with higher similarity in the detection results by rearranging the order of related images, so as to accurately realize the copy detection of local information of an image. Compared with A Self-Supervised Descriptor for Image Copy Detection, this paper achieves a significant improvement on the DISC2021 dataset, with 10% increase in localized copy image recognition accuracy, and 10% - 15% increase in μAP (Micro Average Precision) metrics. In addition, this paper also carries out an intuitive qualitative analysis using heat maps, which confirms that ViT is able to acutely capture the region of the image where local copies exist, further validating the effectiveness of the algorithm. The ViT-based localized image copy detection reordering algorithm proposed in this paper is able to effectively detect copy images using picture-in-picture image enhancement, and achieves excellent results on the DISC2021 dataset, which expands new ideas for research in the field of digital image forensics. |
| Key words: image copy detection ViT image enhancement reordering heat map |
