AGU2-Net: Multi-Scale U2-Net Enhanced by Attention Gate Mechanism for Image Tampering Localization

The advancement of image editing and compositing technologies has posed significant challenges to the authenticity of digital images. Although deep learning algorithms based on convolutional neural networks (CNNs) have made notable progress in image forgery detection, they still face certain limitations in effectively detecting and localizing tampered areas due to the subtle nature of existing manipulation traces. To address this limitation, we propose AGU2-Net, a novel network architecture that integrates an advanced attention mechanism with multi-scale feature extraction. The architecture employs a nested U-structure as its backbone, enabling effective integration of image features across various scales. Additionally, AGU2-Net is an end-to-end autonomous training network with a lightweight framework that incurs low computational overhead. A significant innovation of the architecture is the introduction of the attention gate mechanism, which enhances the network’s ability to selectively process multi-scale features. By adaptively fusing feature information from the encoder and decoder, this mechanism increases spatial perception and improves the network’s ability to capture intricate local tampering details. Experiments on four publicly available benchmarks show that AGU2-Net achieves top F1 scores of 0.556, 0.338, and 0.273 on the CASIA, NIST16, and IMD2020 datasets, respectively, confirming its superior tamper-detection accuracy and sharper pixel-level localization compared with state-of-the-art methods.