An efficient edge detection network inspired by offset receptive fields and feedback modulation

Edge detection is a fundamental task in computer vision, but recent deep learning and bio-inspired methods often suffer from slow inference speeds due to large parameter sizes or complex structural designs, limiting their use in resource-constrained applications. To avoid the computational burden of complex structural designs, abstracting biological visual mechanisms provides a promising alternative. Recent neuroscience studies highlight two efficient mechanisms: (1) offset excitatory–inhibitory receptive fields are highly advantageous for precise edge localization, and (2) top-down feedback dynamically modulates edge integration according to local spatial context. Inspired by these mechanisms, this study proposes a bio-inspired artificial intelligence model for efficient edge detection. First, we design an offset receptive field convolution (ORFC) to simulate the antagonistic response pattern of offset receptive fields. Second, we propose a feedback modulation module (FMM) to model the modulation of shallow edge responses by deeper structural information. By integrating these two components, we construct an Efficient Edge Detection Network (EEDNet). Experiments on three public datasets demonstrate that the proposed method achieves a favorable balance between accuracy and computational efficiency. Specifically, on the Berkeley Segmentation Dataset 500 (BSDS500), EEDNet achieves F-measure of 0.810 at the optimal dataset scale (ODS) and 0.824 at the optimal image scale (OIS), with only 0.62 million parameters, 4.1 billion floating-point operations, and 54.1 frames per second. These results indicate that EEDNet is effective and efficient for resource-constrained edge detection applications.