基于改进ManTra-Net网络的图像篡改检测

Image Tampering Detection Based on Improved ManTra-Net Network

篡改后的图像经常被用于恶意的谣言,威胁社会的稳定性,因此对篡改图像进行检测有利于维持社会信息的准确性。当前,篡改检测技术已经取得了重大进展,但精确识别和定位被篡改的图像区域依旧是一项极具挑战的工作。传统的篡改检测方法只针对某一特定的篡改类型,难以同时针对多种类型的篡改,普遍精度不高,并且检测定位不准确。因而本文提出了一种基于ManTraNet网络的图像篡改检测方法,首先利用ManTra-Net局部异常模块来获取图像篡改信息,使得对图像检测时有聚焦点,其次利用注意力机制来提高对篡改信息的关注,忽略无关的语义信息,增强模型的学习能力,并且基于Unet提出一种新的提取特征网络,提取更加有效且空间信息完整的特征。与RGB-N、SPAN、NOI1、RCRCNN、ManTra-Net五个模型进行比较,实验结果显示本文提出的算法具有相对较高的检测和定位精度。

Tampered images are often used in malicious rumors that threaten the stability of society,so detection of tampered images is beneficial to maintain the accuracy of social information.Tampering detection algorithms exist,and accurately detecting and locating image tampered regions is still a challenging task at present.Traditional tampering detection methods only target a specific type of tampering,making it difficult to detect multiple types of tampering at the same time,generally with low accuracy,and inaccurate detection and localization.Thus,this paper proposes a ManTraNet network-based image tampering detection method,firstly,using the ManTra-Net local anomaly module to obtain image tampering information,which makes a focus on image detection,secondly,using the attention mechanism to improve the focus on tampering information,ignoring irrelevant semantic information,enhancing the learning ability of the model,and proposing a new Unet-based network for extracting features,which extracts more effective and spatially informative features.Comparing with five models,RGB-N,SPAN,NOIl,RCR-CNN,and ManTra-Net,the experimental results show that the algorithm proposed in this paper has relatively high detection and localization accuracy.