RESEARCH RECORD / FG-MoE
* Equal contribution.
METHOD / FINDINGS
Specialized edge, wavelet, texture, and directional filters form an expert ensemble that detects subtle forgery traces in remote-sensing imagery.
A multimodal gating network routes visual and text features to specialized filter-based experts plus a standard expert. Tamper-mask reconstruction, visual question answering, and expert-balancing losses train the model jointly.
92.95% overall accuracy on Real-RSCM. Remote-sensing forgery question answering on the Real-RSCM test set. (Table 1, p. 4.)
A lightweight PCB defect detector that preserves shallow high-resolution features through multi-scale adaptation, semantic alignment, and wavelet enhancement.
Specialized edge, wavelet, texture, and directional filters form an expert ensemble that detects subtle forgery traces in remote-sensing imagery.
A zero-shot template-matching framework that combines visual foundation models, aspect-ratio-preserving normalization, cross-attention, and geometric refinement across image scales.
