HED Edge Detection

Holistically-Nested Edge Detection (HED) is a deep learning-based edge detector developed by Saining Xie and Zhuowen Tu that learns multi-scale hierarchical image boundaries. Unlike traditional threshold operators, HED generates soft, continuous grayscale lines that capture organic curves and volume.

HED outputs soft, anti-aliased grayscale gradients with varying stroke weights, whereas Canny produces binary, single-pixel lines based on strict intensity thresholds. HED is ideal for organic shapes like humans, animals, and flowing drapery because it gives generative models flexibility to restyle surfaces, while Canny is better suited for rigid geometric drafts.

Safe Mode suppresses chaotic micro-textures and high-frequency background noise to produce clean, continuous contours, making it the preferred choice for standard soft-edge conditioning. In contrast, Scribble Mode applies aggressive augmentation to create heavy, coarse outlines resembling hand-drawn marker or charcoal sketches for looser creative reinterpretations.

Avoid HED when you require mathematical precision, pixel-perfect alignment, or crisp straight lines, such as in architectural blueprints, CAD schematics, or typographic extraction. Because HED outputs soft gradients, downstream models may introduce subtle warping to straight edges.

High-contrast images with clearly defined subjects yield the cleanest edge maps. Low-resolution or heavily compressed images should be avoided, as compression artifacts can be mistakenly processed as soft edges.