DiffusionEdge is a novel diffusion-based edge detection model designed to address the challenges of correctness, crispness, and efficiency in edge detection tasks. The model leverages the diffusion probabilistic model (DPM) to directly predict accurate and crisp edge maps without post-processing, which is a significant improvement over traditional and deep learning-based methods. Key contributions include: 1. **DiffusionEdge Model**: The first diffusion model specifically designed for edge detection, which can generate crisp and accurate edge maps. 2. **Technical Designs**: - **Decoupled Architecture**: To speed up the denoising process, a decoupled diffusion architecture is used. - **Adaptive Fourier Filter**: This filter adjusts the latent features of specific frequencies to improve edge detection accuracy. - **Uncertainty Distillation**: The model optimizes the latent space using a distillation strategy, preserving pixel-level uncertainty information. 3. **Performance**: - **Experiments on Datasets**: Extensive experiments on four edge detection benchmarks (BSDS, NYUDv2, Multicue, BIPED) demonstrate superior performance in both correctness and crispness. - **Quantitative Results**: On the NYUDv2 dataset, compared to the second best method, DiffusionEdge increases ODS, OIS (without post-processing), and AC by 30.2%, 28.1%, and 65.1%, respectively. 4. **Conclusion**: - DiffusionEdge shows great potential for downstream tasks due to its ability to generate crisp and accurate edge maps. - Future work will focus on improving the efficiency of the diffusion model to reduce inference time. The paper provides a comprehensive overview of the method, including its architecture, technical details, and experimental results, highlighting its effectiveness and advantages over existing edge detection methods.DiffusionEdge is a novel diffusion-based edge detection model designed to address the challenges of correctness, crispness, and efficiency in edge detection tasks. The model leverages the diffusion probabilistic model (DPM) to directly predict accurate and crisp edge maps without post-processing, which is a significant improvement over traditional and deep learning-based methods. Key contributions include: 1. **DiffusionEdge Model**: The first diffusion model specifically designed for edge detection, which can generate crisp and accurate edge maps. 2. **Technical Designs**: - **Decoupled Architecture**: To speed up the denoising process, a decoupled diffusion architecture is used. - **Adaptive Fourier Filter**: This filter adjusts the latent features of specific frequencies to improve edge detection accuracy. - **Uncertainty Distillation**: The model optimizes the latent space using a distillation strategy, preserving pixel-level uncertainty information. 3. **Performance**: - **Experiments on Datasets**: Extensive experiments on four edge detection benchmarks (BSDS, NYUDv2, Multicue, BIPED) demonstrate superior performance in both correctness and crispness. - **Quantitative Results**: On the NYUDv2 dataset, compared to the second best method, DiffusionEdge increases ODS, OIS (without post-processing), and AC by 30.2%, 28.1%, and 65.1%, respectively. 4. **Conclusion**: - DiffusionEdge shows great potential for downstream tasks due to its ability to generate crisp and accurate edge maps. - Future work will focus on improving the efficiency of the diffusion model to reduce inference time. The paper provides a comprehensive overview of the method, including its architecture, technical details, and experimental results, highlighting its effectiveness and advantages over existing edge detection methods.