Prompting and Tuning: In-Band Interference Segmentation Using Segment Anything Model

This letter explores potential of a segment anything model (SAM), the first promptable image segmentation system, in detecting wireless interference based on time-frequency images (TFI). Since the original SAM is pre-trained with natural images, it struggles to segment interferences within bandwidth of legitimate communication signals. To address this, we propose a novel prompting and tuning enhanced SAM to effectively segment in-band interferences, especially in the presence of blurred boundaries on TFIs. We first exploit the difference in single-channel luminance among gray-scaled TFI pixels and propose a mean shift clustering on this basis which helps to prompt the interference regions for segmentation. Then the lightweight mask decoder of SAM is fine-tuned using an augmented dataset containing the refined examples of in-band interferences. Compared to the state-of-the-art, experiments show that the proposed method significantly improves the segmentation accuracy as indicated by dice coefficients and intersection-over-union in a wide rang of jamming-to-signal ratio.