Fast Detection of Curved Edges at Low SNR

CVPR 2016 低信号噪声比图像中曲线边缘快速检测

http://www.wisdom.weizmann.ac.il/~yehonato/projectPage.html

有源代码

本文针对低信号噪声比图像 曲线边缘检测主要思路如下：
1）An efficient hierarchical algorithm to examine an exponential number of candidate curved edges 快速的将可能是曲线边缘的地方检测出来

Examine each potential edge curve using its “custom tailored” matched filter

Do this efficiently using a dynamic programming-like algorithm on a hierarchical, binary-split tree of the image (keep best curve for each two points on the boundary of a tile)

2） 抑制噪声： Long is good: noise can be averaged out by smoothing along the curve (while maintaining contrast across the curve) by using a matched filter
对边缘两侧进行平滑，去除噪声

3） Use statistically rigorous adaptive threshold to detect edges at very
low SNRs 使用自适应阈值检测边缘

Computational complexity:
Stringent: o(N1.5)
examine all contact points in the interface between tiles

Greedy: O（Nlog N）
contact points are sorted by score; only curves through highest scoring points are examined

Runtime:0.9 (0.6) secs on 129×129