百科.dev
全部条目在线工具AI 编程趋势榜开源项目技术资讯提交条目
登录
< 返回工具列表
C

CurvatureFilter

> 编程语言
开源

曲率过滤器是变分模型的高效求解器

394 stars0 点赞0 次浏览
访问官网GitHub

工具介绍

曲率过滤器是变分模型的高效求解器

Curvature filters are efficient solvers for variational models.

These curvature filters are developed by Yuanhao Gong during his PhD. MC filter and TV filter are exactly the same as described in the paper. But the GC filter is slightly modified. Please cite following papers if you use curvature filter in your work. Thank you!

:books: The Paper. The general theory is in Chapter Six of PhD thesis (downloaded 17,000+)

:closed_book: Presentation of Gaussian Curvature Filter: LinkedIn, Dropbox or Baidu.

:blue_book: Poster of Bernstein Filter can be found here.

:gift: a short introduction in Chinese (中文): Zhihu(Editors' Choice) or this Zhihu

:trophy: source code in C++ and Java can also be found from MOSAIC group

:bell: The kernels summary and one example how to get the kernel can be found here.

:e-mail: [email protected] or join the Curvature Filter Forum


…

Curvature filters' philosophy

Traditional solvers, such as gradient descent or Euler Lagrange Euqation, start at the total energy and use diffusion scheme to carry out the minimization. When the initial condition is the original image, the data fitting energy always increases while the regularization energy always reduces during the optimization, as illustrated in the below figure. Thus, regularization energy must be the dominant part since the total energy has to decrease.

Therefore, Curvature filters focus on minimizing the regularization term, whose minimizers are already known. For example, if the regularization is Gaussian curvature, the developable surfaces minimize this energy. Therefore, in curvature filter, developable surfaces are used to approximate the data. As long as the decreased amount in the regularization part is larger than the increased amount in the data fitting energy, the total energy is reduced.


Features

Theoretical Practical
Generality: handle arbitrary data fitting term (BlackBox) Efficient: three or four order of magnitude faster than traditional solvers
Convergence: theoretically guaranteed Implementation: 40 lines in Matlab and 100 lines in C++

Faster and Faster

Filter Bilateral Filter Guided Filter Guided Filter MC Filter GC Filter Bernstein Filter
Lang. C++ Matlab C++ C++ C++ C++
MilliSec. 103 514 130 8 (or 327 MPixels/sec) 11 7

Running time with 10 iterations on 512X512 Lena image. Matlab version is R2015a and GCC version is 5.1. All tests are on a Thinkpad T410 with i7-620M core CPU (2.6GHz). We take the time for 100 iterations and divide it by 10. On average, curvature filters take 1 millisecond per iteration.

On my new taptop(Thinkpad T470p, NVIDIA GeForce 940MX, 384 CUDA cores), GPU version of MC filter can achieve 2500 MPixels/Second with shared memory and single precision.

On the TITAN Xp card, MC filter can achieve 33.2 Giga Pixels/Second with shared memory and single precision. On the Tesla K40c card (2880 cores), MC filter can achieve 8090 MPixels/Second with shared memory and single precision.


Example Applications

1) Denoising

The noise free test image can be downloaded here

2) Only minimize the regularization

GC = Gaussian Curvature, MC = Mean Curvature, TV = Total Variation

3) Minimize a variational model, showing the line profile

We show three lines' profiles during minimizing a mean curvature regularized model (MC filter used).

4) Cartoon Texture Decomposition

5) Registration

from left to right: original reference image, distorted source image, registered results by TV filter, MC filter and GC filter.


On Triangular Meshes (preliminary results, p.195 in the thesis)

original mesh (left) and processed mesh (right), the energy profile is shown in the middle.


FAQ:

  1. Why dual mesh (DM) structure is needed?

There are two reasons. First, these four sets guarantee the convergence. Second, we can use the updated neighbors for current position. Therefore, it is more computational efficient.

  1. What is the difference between these three filters?

In general, GC filter is better in preserving details, compared with the other two. And TV filter is better in removing noise as well as details. MC filter is between these two.

These three filters are correspond to three types of variational models. User should decide which prior is to be assumed about the ground truth.

  1. What is the difference between split and nosplit scheme?

In general, splitting the image into four sets and looping on them is computational faster. However, in some cases like deconvolution, we need to merge the four sets after every iteration. So, it is better do nosplit scheme.

These two lead to exactly the same result. The split code is just more cache friendly.

Issues· 0 开放

查看全部 Issues在 GitHub 打开

暂无开放 Issues,或尚未同步最近议题。

> 标签

C++curvaturecurvature-filtersdifferential-geometryimage-processing

暂无评论,来聊聊你的看法吧

> 工具信息

发布日期2026年8月1日
最后更新2026年9月17日
分类编程语言
定价开源

> 相关工具

T
TypeScript
JavaScript 的超集,为前端与全栈提供静态类型
P
Python
通用编程语言,广泛用于 Web、数据与 AI
G
Go
Google 推出的简洁高效系统语言