dev.to

DiffusionEdge: Diffusion Probabilistic Model for Crisp Edge Detection

Yunfan Yeが第一著者，Zhiping Caiグループ，防衛系の組織？ In this paper, the authors proposed DiffusionEdge, which can detect edges of natural images. DiffusionEdge includes a learnable filter in the process of the Fourier transform. In addition, the proposed method calculates back propagation more efficiently by omitting components that appear in the process. This is the first paper that applies Diffusion to edge detection. Related works Diffusion models Edge detection Proposed method Loss is calculated in an adaptive way as follows: lij=αlog(1−pij)if  Ei=0 l_i^j= \alpha log(1-p_i^j) \quad {\rm if} \; E_i = 0 lij​=αlog(1−pij​)ifEi​=0 lij=0if  Ei∈(0,η) l_i^j= 0 \quad {\rm if} \; E_i \in (0, \eta) lij​=0ifEi​∈(0,η) lij=βlogEijif  o.w l_i^j= \beta logE_i^j \quad {\rm if}\; o.w lij​=βlogEij​ifo.w where p_i^j means edge prbability for the i-th pixel in the j-th edge. \alpha and \beta are adaptively tuned by the ground truth of the reference, respectively. Lwce=∑lij L_{wce} = \sum l_i^j Lwce​=∑lij​ Also, omission is conducted as follows: ParseError: KaTeX parse error: Undefined control sequence: \nable at position 2: \̲n̲a̲b̲l̲e̲_\theta L_{wce}… Actually, the mid-term is omitted for the efficiency. 感想 本来ある数学的操作を飛ばすことも工学的には重要なこともあるのだなあ そういう意味で物理学的なのかもしれないCSは

May 13, 2025 - 01:26
 0
DiffusionEdge: Diffusion Probabilistic Model for Crisp Edge Detection

Yunfan Yeが第一著者，Zhiping Caiグループ，防衛系の組織？

In this paper, the authors proposed DiffusionEdge, which can detect edges of natural images.
DiffusionEdge includes a learnable filter in the process of the Fourier transform.
In addition, the proposed method calculates back propagation more efficiently by omitting components that appear in the process.
This is the first paper that applies Diffusion to edge detection.

Related works
Diffusion models
Edge detection

Proposed method
Loss is calculated in an adaptive way as follows:

lij=αlog(1−pij)if  Ei=0 l_i^j= \alpha log(1-p_i^j) \quad {\rm if} \; E_i = 0 lij​=αlog(1−pij​)ifEi​=0
lij=0if  Ei∈(0,η) l_i^j= 0 \quad {\rm if} \; E_i \in (0, \eta) lij​=0ifEi​∈(0,η)
lij=βlogEijif  o.w l_i^j= \beta logE_i^j \quad {\rm if}\; o.w lij​=βlogEij​ifo.w

where p_i^j means edge prbability for the i-th pixel in the j-th edge.

\alpha and \beta are adaptively tuned by the ground truth of the reference, respectively.

Lwce=∑lij L_{wce} = \sum l_i^j Lwce​=∑lij​

Also, omission is conducted as follows:

ParseError: KaTeX parse error: Undefined control sequence: \nable at position 2: \̲n̲a̲b̲l̲e̲_\theta L_{wce}…

Actually, the mid-term is omitted for the efficiency.

感想
本来ある数学的操作を飛ばすことも工学的には重要なこともあるのだなあ
そういう意味で物理学的なのかもしれないCSは

Read More

Tags:

Previous Article

From Chaos to Clarity: Making Your Data AI-Ready

Next Article

Unlocking Creativity: How an AI HTML Code Generator Can Transform Your Web Devel...

Related Posts

Experience with Fellou: The World’s First Agentic Browser

Experience with Fellou: The World’s First Agentic Browser

May 12, 2025  0

May 11, 2025  0

3D Wireframes Evolved: Continuous Latent Model Creates Smooth, Complex Structures

3D Wireframes Evolved: Continuous Latent Model Creates ...

May 5, 2025  0