没有合适的资源?快使用搜索试试~ 我知道了~
首页基于CPDA的鲁棒图像角点检测技术
本文主要探讨了"基于Chord-to-Point Distance Accumulation (CPDA)技术的鲁棒图像角点检测方法"。该研究发表在2008年10月的IEEE TRANSACTIONS ON MULTIMEDIA上,由Mohammad Awrangjeb和Guojun Lu两位作者提出,他们是IEEE的成员和高级成员。 CSS(曲率尺度空间)是许多基于轮廓的图像角点检测算法的基础。然而,CSS方法存在一些局限性。首先,CSS中的“曲率”概念对于局部变化和噪声非常敏感,除非在计算前进行适当的平滑处理。这可能导致结果的不稳定性和误差,特别是在高阶导数(如二阶导数)的计算中。这在实际应用中可能会影响角点检测的准确性。 CSS方法的另一个问题是,高斯滤波会对曲线产生影响,选择合适的平滑尺度变得困难,从而影响CSS角点检测技术的整体性能。为了克服这些问题,作者提出了一个全新的角点检测技术——CPDA,专注于离散曲率的估计。 CPDA技术的优势在于它对曲线上的局部变化和噪声具有较低的敏感度。相比于依赖CSS中的高阶导数,CPDA通过累积弦到点的距离,提供了一种更为稳健的方法来评估图像中的角点特征。这种方法减少了由于平滑操作引入的复杂性,并提高了检测结果的稳定性,尤其是在处理噪声或图像细节时。 总结来说,这篇论文的核心贡献在于提出了一种创新的角点检测策略,它能够更有效地应对图像中的噪声和复杂性,从而提高角点定位的准确性和鲁棒性。这对于计算机视觉、图像处理和机器学习等领域中的许多应用具有重要的实际价值。如果你对这个主题感兴趣,作者的博客提供了深入的讨论和交流机会,可以进一步探索和学习CPDA技术的细节和应用实例。
资源详情
资源推荐
AWRANGJEB AND LU: ROBUST IMAGE CORNER DETECTION BASED ON THE CHORD-TO-POINT DISTANCE ACCUMULATION TECHNIQUE 1061
Fig. 3. Effect of curve smoothing using Gaussian function with different
smoothing-scale
. The original curve as shown above is a high resolution
version of “curve 7” in Fig. 5(b).
when the curve point-locations are always approximated and,
therefore, may not be stable. Consequently, the derivative-based
curvature value defined in (4) may change considerably between
original and transformed curves. Such unstable and erroneous
estimated curvature not only affects the corner detection perfor-
mance under geometric transformations, but also may result in
poor corner matching performance in any of their later applica-
tions [4].
The second problem is related to the curve smoothing using
the Gaussian function. The aim of the smoothing is to reduce
the effect of local variation and noise so as to remove weak
and false corners. However, two direct effects of the Gaussian
smoothing are: first, it shrinks the curve and second, it smoothes
out the details if the smoothing-scale is high. On a smoothed
curve, corners are usually detected at those points where the
absolute curvature maxima values become higher than the
threshold. However, the estimated curvature value at a corner
point may decrease below the threshold with smoothing using
high
. The sharp (strong) corners are easily identifiable with
smoothing using large
, their estimated curvature is less accu-
rate and localization is worse than with smoothing using small
. However, using small may detect many weak and false
corners. Moreover, one smoothing-scale may not be suitable
for all curves. Since the local variation and noise on the curve
are unknown, curves of the same length may require different
smoothing-scales, even different segments of the same curve
may require different smoothing-scales. Therefore, choosing
an appropriate Gaussian smoothing-scale for a given curve is
a very difficult task.
This above problem is depicted in Fig. 3. The original curve
in Fig. 3 is a high resolution version of “curve 7” in Fig. 5(b),
where we see three strong corners (points 2, 3, and 4 in Fig. 3). In
the high resolution version of the curve (original curve in Fig. 3),
we see many weak corners and we choose points 1, 5, and 6
for illustration. Smoothing with both
and 4 can detect all
three strong corners; but while smoothing with
detects
all the weak corners, smoothing with
misses two weak
corners (points 1 and 5). However, corner localization is better
with lower smoothing-scale (see locations of point 3 in original
and smoothed curves in Fig. 3). Note that the detected corners
are shown with solid arrows and the missed corners are shown
with dotted arrows.
2) Existing CSS-Based Detectors: The first problem dis-
cussed above was an inherent problem with all the existing CSS
corner detectors [3], [5]–[11], since they used the same CSS
curvature estimation technique discussed in Section II-A. In the
rest of this section, we will discuss how the existing CSS-based
detectors suffer from the second problem.
The earliest CSS corner detector by Rattarangsi and Chin [5]
and its improved version [6] detected corners in the complete
scale-space by considering all possible smoothing-scales so as
to overcome the second problem. However, the tree construc-
tion and parsing using the complete scale-space resulted in high
computational complexity. The adaptive smoothing technique
in [7] reduced the computational complexity by eliminating the
construction of the complete scale-space map and by avoiding
the tree representation of the scale-space. However, the adaptive
smoothing technique itself incurred additional computational
cost which was supposed to be compensated by the slightly im-
proved robustness.
Mokhtarian and Suomela [8] detected corners at a high scale
and tracked them through multiple lower scales to the lowest
scale in order to improve localization. On the one hand, when
corners were detected in a very high scale, this detector showed
high robustness but lost many strong corners which might re-
quire lower smoothing-scales to be detected. On the other hand,
if corners were detected in a low scale it introduced many weak
corners which might require higher smoothing-scales to be re-
moved. As a result, this detector also suffered from the second
problem. Many of its improved versions (e.g., [3], [9]), which
selected the smoothing-scales based on the curve-length, also
failed to overcome this problem. Because even curves of the
same length may require different smoothing-scales depending
on the level of local variation and noise as discussed above.
The multi-scale curvature product (MSCP) detector [10]
and the single-scale detector [11] tried to overcome the second
problem to a great extent. They compensated the risk asso-
ciated with possible inappropriate smoothing-scale selection
by adopting different strategies. The MSCP detector used the
curvature product of three estimated curvature values at each
point using three smoothing-scales. As a result, in terms of cur-
vature product, the strong corners became more distinguishable
from the weak corners. He and Yung [11] used the adaptive
curvature-threshold and the dynamic region-of-support on both
sides of each curvature extremum point. Consequently, both of
them offered better corner detection performance than many of
the aforementioned detectors. In this paper, we will compare
most promising detectors [3], [8], [10], [11] with our proposed
corner detector (see Section IV).
C. CPDA Discrete Curvature Estimation
The previously proposed single chord-to-point distance mea-
surement technique [14] was not reliable, since such distance
depends upon the location of the chord [12]. In order to increase
the reliability, Han and Poston [12] proposed the chord-to-point
distance accumulation (CPDA) technique for measuring the dis-
crete curvature. A chord is moved along a curve and the perpen-
dicular distances from each point on the curve to the chord are
summed to represent the curvature. In contrast to the conven-
tional CSS curvature estimation technique adopted by the ex-
isting CSS corner detectors, the CPDA technique is completely
based on the Euclidean distance and does not involve any deriva-
tive of the curve-point locations at all.
剩余13页未读,继续阅读
tostq
- 粉丝: 1350
- 资源: 20
上传资源 快速赚钱
- 我的内容管理 展开
- 我的资源 快来上传第一个资源
- 我的收益 登录查看自己的收益
- 我的积分 登录查看自己的积分
- 我的C币 登录后查看C币余额
- 我的收藏
- 我的下载
- 下载帮助
最新资源
- C++标准程序库:权威指南
- Java解惑:奇数判断误区与改进方法
- C++编程必读:20种设计模式详解与实战
- LM3S8962微控制器数据手册
- 51单片机C语言实战教程:从入门到精通
- Spring3.0权威指南:JavaEE6实战
- Win32多线程程序设计详解
- Lucene2.9.1开发全攻略:从环境配置到索引创建
- 内存虚拟硬盘技术:提升电脑速度的秘密武器
- Java操作数据库:保存与显示图片到数据库及页面
- ISO14001:2004环境管理体系要求详解
- ShopExV4.8二次开发详解
- 企业形象与产品推广一站式网站建设技术方案揭秘
- Shopex二次开发:触发器与控制器重定向技术详解
- FPGA开发实战指南:创新设计与进阶技巧
- ShopExV4.8二次开发入门:解决升级问题与功能扩展
资源上传下载、课程学习等过程中有任何疑问或建议,欢迎提出宝贵意见哦~我们会及时处理!
点击此处反馈
安全验证
文档复制为VIP权益,开通VIP直接复制
信息提交成功