BASNet: Boundary Assisted Network for Image Splicing Forgery Detection

Image splicing is a common technique used in image forgery. With the rapid development of digital image processing technology, detecting image splicing forgery has become increasingly challenging. Existing splicing forgery localization methods lack exploration in effectively utilizing tampered region boundary information. To address this issue, we propose a novel model for detecting image splicing forgery called boundary-assisted network (BASNet). We introduce a boundary-motivated module (BMM) to explore valuable and additional boundary features related to tampered regions, enhancing representation learning for detecting tampered regions. Additionally, we present a boundary-enhanced module (BEM) to enhance boundary information using the cross-channel attention mechanism. To efficiently merge features from various levels and boundary features, we further present the feature fusion module (FFM). To optimize performance, the BASNet incorporates weighted binary cross-entropy loss, dice loss, and boundary loss, which can effectively leverage edge supervision while mitigating imbalance between positive and negative samples. Evaluation of five widely-used forgery detection datasets demonstrates the state-of-the-art performance of the BASNet. Robustness experiments verify that the BASNet is robust enough to detect image splicing forgery across various common attacks.