Frequency-Domain Decomposition and Recomposition for Robust Audio-Visual Segmentation
The Frequency-Aware Audio-Visual Segmentation (FAVS) framework reformulates audio-visual segmentation as a frequency decomposition and recomposition problem to resolve contradictions between audio noise and visual detail. It introduces two modules, FDED for modality-specific feature enhancement and SCMC for cross-modal consistency using a mixture-of-experts architecture. The method achieves state-of-the-art performance on benchmark datasets by isolating and synergistically integrating frequency bands.