Multiscale feature fusion and local detail enhancement-based edge detection

Existing edge detection methods often fail to balance global structure and local details in scenarios with complex backgrounds and rich details. In this paper, we proposed a multi-scale feature fusion and local detail enhancement based edge detection method ML-EdgeNet. ML-EdgeNet is constructed based on Bi-Directional Cascade Network (BDCN). Subsequently, we intorduced Centralized Feature Pyramid (CFP) to capture multi-level information and avoid blurry or fragmented edges. Finally, a Multi-Dilated Channel Refiner Module (MDCRM) is proposed to refine edge regions with multi-scale dilated depthwise convolution. Comprehensive experiments were conducted on the BSDS500 dataset. Experimental results show that ML-EdgeNet achieves significant performance improvement. Compared with existing edge detection methods, ML-EdgeNet has achieves more than 0.023 Optimal Dataset Scale with F-measure (ODS-F) improvement, which demonstrates its effectiveness and robustness in practical applications.