DenseEdgeNet: A Lightweight Deep Edge Detection Network With Hierarchical Supervision and Generalization Across Visual Domains

Existing edge detectors often depend on large pretrained backbones and still miss delicate, fine-scale contours. We introduce DenseEdgeNet, a compact architecture that performs hierarchical multi-scale feature extraction and employs a dual-path upsampling strategy. Dense skip links retain early edge cues, while an attention-guided fusion block enforces structural coherence across scales. The dual upsampling branch explicitly negotiates the trade-off between semantic abstraction and spatial detail. Training uses an enhanced IoU-oriented objective, IoULoss+, which increases responsiveness to weak, low-contrast edges. Without any pretrained initialization, DenseEdgeNet attains accurate and broadly generalizable edge maps, surpassing prior methods on BIPED and MDBD under similar parameter budgets and demonstrating stronger cross-dataset transfer.