Coarse-to-fine spatial-channel-boundary attention network for image copy-move forgery detection
Jun-Liu Zhong,Ji-Xiang Yang,Yan-Fen Gan,Lian Huang,Hua Zeng
DOI: https://doi.org/10.1007/s00500-022-07432-x
IF: 3.732
2022-10-01
Soft Computing
Abstract:Forensics still faces a serious challenge with image copy-move forgery, in which the copied source and pasted target regions exist in the same image, also known as a homogeneous forgery. In the past decade, numerous copy-move forgery detection (CMFD) methods have attempted to resolve this issue. However, the traditional keypoint-based and block-based methods have certain insurmountable deficiencies, such as the inability to smooth out regions and the lack of scaling invariance. Since the introduction of deep neural networks (DNNs) in the CMFD scheme, researchers have been able to overcome the defects of the traditional hand-crafted methods and obtain promising results. Using DNNs as a reference, this paper proposes a coarse-to-fine spatial-channel-boundary attention network (SCBAN) which is more suited to CMFD. SCBAN consists of three sub-networks, namely, feature extraction, coarse forgery identification, and fine forgery identification modules. First, CondenseNet will serve as SCBAN's backbone for feature extraction. Next, we present a dual-correlation-attention module for parallel fusion, as well as a nearest-correlation matching module for coarse forgery identification. In addition, we propose a boundary refinement attention module for fine forgery identification. We have conducted numerous experiments on IMD, CoMoFoD, and CMHD benchmarks to show that our SCBAN can achieve the best performance and robustness, compared to the existing DNN CMFD. In addition, unlike the well-designed hand-crafted methods which achieve good performance in a specific dataset, our SCBAN can maintain its scalability to achieve good performance on multiple benchmark datasets.
computer science, artificial intelligence, interdisciplinary applications