Answers are grounded in this paper and cite their sources.
In this paper, we first propose a dual-stream multi-scale feature fusion image tampering localization algorithm based on the attention mechanism, where the noise branch utilizes SRM filter to suppress the interference of the image content information, and extracts additional tampering clues from the noise domain information. In addition roaster advanced semantic information as well as tamper edge feature information in RGB color domain simultaneously. The ELA feature information of the image is extracted using ELA extractor, followed by inputting the three sets of feature information into the model for effective fusion. Finally, a tamper localization method based on hierarchical contrast learning is proposed by introducing contrast learning and the lesser loss function. For the pre-training experiments, this paper is much better than ManTraNet and MVSS-Net in terms of AUC and F 1 . For the fine-tuning experiments, this paper’s method has a better performance advantage than MVSS-Net in terms of copypaste and splicing operations, and the number of parameters of 7.84M is much smaller than the number of parameters of MVSS-Net of 145.61M. It is verified that this paper’s method can maintain a high detection performance and reduce the computation volume at the same time. It is verified that the method in this paper can maintain high detection performance while reducing the computation amount.
No takes yet. Share an insight, caveat, or question.
Qi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: