TY - GEN
T1 - DSTNet
T2 - 27th International Conference on Pattern Recognition, ICPR 2024
AU - Zhao, Kaiqi
AU - Yuan, Xiaochen
AU - Huang, Guoheng
AU - Liu, Kun
N1 - Publisher Copyright:
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.
PY - 2025
Y1 - 2025
N2 - In copy-move forgery detection, most relevant studies concern locating the copy-move areas without the distinction of source and target regions. This paper proposes an end-to-end network, DSTNet, to identify the source and target based on consistency detection between the copy-move region and the non-copy-move region. The DSTNet is composed of two stages, the Pre-processing stage and the Discrimination stage. Pre-processing Stage extracts internal information of copy-move and non-copy-move areas and conducts a series of operations to meet the requirements of network input. Discrimination stage allows multiple patches for input and classifies the input patches. Specifically, the Pre-processing stage, contains the Copy-move Patches Selection (CM Patches Selection) and Genuine Patches Selection, can select pairs of copy-move and none copy-move patches. We train the proposed DSTNet on two large synthetic datasets and use the public datasets CASIA and Comofod for evaluation. The experiment shows that our method achieves excellent results. Particularly, we achieve a 5.4% higher F1 based on ground-truth of copy-move mask (GT-CM) on CASIA dataset.
AB - In copy-move forgery detection, most relevant studies concern locating the copy-move areas without the distinction of source and target regions. This paper proposes an end-to-end network, DSTNet, to identify the source and target based on consistency detection between the copy-move region and the non-copy-move region. The DSTNet is composed of two stages, the Pre-processing stage and the Discrimination stage. Pre-processing Stage extracts internal information of copy-move and non-copy-move areas and conducts a series of operations to meet the requirements of network input. Discrimination stage allows multiple patches for input and classifies the input patches. Specifically, the Pre-processing stage, contains the Copy-move Patches Selection (CM Patches Selection) and Genuine Patches Selection, can select pairs of copy-move and none copy-move patches. We train the proposed DSTNet on two large synthetic datasets and use the public datasets CASIA and Comofod for evaluation. The experiment shows that our method achieves excellent results. Particularly, we achieve a 5.4% higher F1 based on ground-truth of copy-move mask (GT-CM) on CASIA dataset.
KW - Copy-move forgery detection
KW - Deep learning for forensics
KW - Siamese network
KW - Source and target areas distinguishing
UR - http://www.scopus.com/inward/record.url?scp=85212256984&partnerID=8YFLogxK
U2 - 10.1007/978-3-031-78312-8_21
DO - 10.1007/978-3-031-78312-8_21
M3 - Conference contribution
AN - SCOPUS:85212256984
SN - 9783031783111
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 321
EP - 333
BT - Pattern Recognition - 27th International Conference, ICPR 2024, Proceedings
A2 - Antonacopoulos, Apostolos
A2 - Chaudhuri, Subhasis
A2 - Chellappa, Rama
A2 - Liu, Cheng-Lin
A2 - Bhattacharya, Saumik
A2 - Pal, Umapada
PB - Springer Science and Business Media Deutschland GmbH
Y2 - 1 December 2024 through 5 December 2024
ER -