• We propose SCIFL-Net, an RGB-noise dual-branch localization network that accurately segments tampered regions at the pixel level. • We design a Wavelet-Enhanced Adaptive Feature Fusion (WEAFF) module to improve the extraction of tampering-sensitive structures. • We develop a Structure-Guided Feature Fusion (SGFF) module to enhance the network’s ability to perceive subtle structural perturbations. Seam carving has emerged as a highly imperceptible form of image tampering. Unlike conventional forgeries, it preserves global semantics while introducing subtle structural distortions that existing localization methods can roughly capture, but often with limited precision, leading to missed detections and false alarms. To address this issue, we propose SCIFL-Net (Seam Carving-Based Image Forgery Localization Network), a deep learning architecture built upon the RRUNet backbone. SCIFL-Net adopts a dual-branch design that jointly exploits noise and semantic features. By incorporating multiple noise-sensitive features, the SE attention mechanism, and Haar wavelet downsampling, the network substantially enhances its ability to capture fine-grained structural disturbances. Moreover, we introduce a Wavelet-Enhanced Adaptive Feature Fusion (WEAFF) module, which strengthens noise features in the frequency domain and subsequently performs adaptive fusion to suppress redundancy while emphasizing tampering-sensitive cues. In addition, the Structure-Guided Feature Fusion (SGFF) module enables deep cross-fusion between noise and semantic representations, progressively guiding the RGB branch to focus on potential tampered regions and thereby improving structural awareness and localization precision. Extensive experiments demonstrate that SCIFL-Net achieves state-of-the-art performance in seam carving localization, enabling accurate and fine-grained detection of tampered regions.
Liu et al. (Mon,) studied this question.