Image-free ghost edge detection at ultra-low sampling rates via a Vision Mamba network

Edge detection is essential for extracting structural boundaries in visual scenes, but conventional deep-learning-based methods remain sensitive to degradations in the input images and involve redundant full-image acquisition. Image-free ghost edge detection provides an efficient alternative by directly extracting edge information from bucket measurements, yet existing methods still require relatively high sampling rates to obtain reliable edge maps. Here, using a Vision Mamba network within the ghost edge detection framework, we achieve image-free edge detection directly from bucket measurements at ultra-low sampling rates. By employing joint optimization of illumination patterns and network parameters, the proposed method achieves reliable edge extraction at a sampling rate as low as 0.39%, as demonstrated by both simulations and optical experiments. Furthermore, the demonstrated robustness to measurement noise and qualitative edge-recovery results for selected visually distinct targets indicate the potential of the proposed method under challenging measurement conditions. We envision this image-free ghost edge-detection framework as a prospective platform for efficient task-oriented imaging under severe measurement constraints.