High-Precision Edge Detection via Task-Adaptive Texture Handling and Ideal-Prior Guidance

Image edge detection (ED) requires specialized architectures, reliable supervision, and rigorous evaluation criteria to ensure accurate localization. In this work, we present a framework for high-precision ED that jointly addresses architectural design, data supervision, and evaluation consistency. We propose SDPED, a compact ED model built upon Cascaded Skipping Density Blocks (CSDB), motivated by a task-adaptive architectural transfer from image super-resolution. By re-engineering texture-oriented structures for ED, SDPED effectively differentiates textures from edges while preserving fine spatial precision. Extensive experiments on four benchmark datasets (BRIND, UDED, MDBD, and BIPED2) demonstrate consistent performance improvements, particularly in Average Precision (AP), with gains of up to 22.5% on MDBD and 11.8% on BIPED2. In addition, we introduce an ideal-prior guidance strategy that incorporates noiseless data into training by treating labels as noise-free samples, providing a practical means to mitigate the subjectivity and noise inherent in human annotations. To enable fair and resolution-independent evaluation, we further adopt a fixed-pixel criterion for assessing localization accuracy. Overall, this work offers a coherent solution for high-precision ED and provides insights applicable to precision-oriented modeling in low-level and soft-computing-based vision tasks. Codes can be found on https://github.com/Hao-B-Shu/SDPED.

💡 Research Summary

This paper tackles two persistent challenges in edge detection (ED): insufficient spatial precision under strict evaluation criteria and the detrimental effect of noisy human annotations. To address these issues, the authors propose a unified framework that combines architectural innovation, a novel training paradigm, and a rigorous evaluation protocol.

Architectural contribution – The core of the proposed model, SDPED (Skipping Density Precise Edge Detection), is built upon Cascaded Skipping Density Blocks (CSDB). CSDB originates from dense blocks used in super‑resolution (SR) networks, which excel at handling high‑frequency texture information. Directly applying a standard dense block to ED, however, leads to blurred edge localization because the dense connectivity overly mixes texture and edge features. The authors therefore redesign the block into a cascaded skipping structure: after each dense sub‑module, a skip connection bypasses intermediate layers, preserving edge‑specific signals while still benefiting from the texture‑suppressing capacity of dense connections. Four CSDB stages form the main backbone, followed by a dedicated fusion block composed of three convolutions (including two 3×3 kernels) and a Leaky ReLU, which aggregates the rich semantic information extracted by the CSDBs. The resulting network contains roughly 1.2 million parameters and requires about 3.4 GFLOPs, making it lightweight enough for real‑time inference on mobile hardware.

Training paradigm – Ideal‑Prior Guidance – Human‑annotated edge maps inevitably contain subjectivity and labeling noise, limiting the ceiling of any supervised ED model. The authors introduce an “ideal‑prior” augmentation strategy that treats the ground‑truth edge map itself as a noise‑free sample. During training, each image is paired not only with its noisy annotation but also with a version of the annotation used as an input (or slightly transformed version) alongside the original image. This forces the network to learn a mapping from a perfect edge prior to the same perfect output, effectively teaching the model what an ideal edge looks like. Experiments show that when the model receives a noiseless edge map as input, it produces almost perfect predictions (average precision ≈ 99.8 %). When trained with this augmentation, SDPED also exhibits markedly sharper predictions on regular images, demonstrating improved discrimination between texture and true edges.

Evaluation protocol – Fixed‑Pixel Criterion – Traditional ED benchmarks evaluate performance using ODS/OIS scores derived from a tolerance radius expressed as a fraction of the image diagonal (maxDist). Because this fraction translates to different absolute pixel tolerances for images of varying resolution, it obscures true localization accuracy. The authors propose a resolution‑independent metric: a fixed‑pixel tolerance (e.g., 2 px). Under this stricter, absolute criterion, SDPED consistently outperforms state‑of‑the‑art methods across four datasets (BRIND, UDED, MDBD, BIPED2). Notably, average precision (AP) improves by up to 22.5 % on MDBD and 11.8 % on BIPED2 compared with the previous best models.

Experimental validation – The paper conducts extensive experiments on the four aforementioned benchmarks, comparing against seven strong baselines (HED, RCF, BDCN, DexiNed, PiDiNet, TEED, and a recent transformer‑based approach). While ODS/OIS scores are comparable to the baselines, the proposed Fixed‑Pixel AP metric reveals a substantial gap in favor of SDPED. Ablation studies confirm that (1) the CSDB backbone contributes the majority of the gain, (2) the ideal‑prior augmentation further boosts performance, especially on noisy datasets, and (3) the dedicated fusion block improves final edge map refinement.

Limitations and future work – The ideal‑prior strategy relies on the availability of clean edge maps, which may not exist in domains lacking precise annotations (e.g., medical imaging). Extending the concept to self‑supervised or synthetic priors is suggested as future work. Additionally, the current CSDB depth (four stages) is fixed; adaptive depth or neural architecture search could yield even more efficient configurations.

Conclusion – By transferring texture‑handling mechanisms from super‑resolution to edge detection, augmenting training with noise‑free priors, and enforcing a resolution‑independent evaluation, the authors deliver a compact yet high‑precision edge detector. SDPED sets a new benchmark for accurate edge localization while remaining computationally lightweight, offering a practical solution for both research and real‑world vision systems that demand fine‑grained edge information.