鲁棒像素顺序特征描述新方法对抗高斯噪声

需积分: 3 21 浏览量更新于2024-10-05 1 收藏 455KB PDF 举报

本文主要探讨了"Robust Order-based Methods for Feature Description"，由Raj Gupta、Harshal Patel和Anurag Mittal三位作者在印度理工学院马德拉斯分校计算机科学与工程系发表。在现代计算机视觉领域，基于特征的方法在物体识别、三维重建和图像拼接等应用中越来越受到重视。传统的诸如SIFT (尺度不变特征变换)、GLOH (全局局部二值模式直方图) 和 Shape Context 等方法，通常依赖于像素强度直方图来描述特征，尽管它们在某些场景下表现良好，但面对高斯噪声时，其鲁棒性有所欠缺。文章提出了两种改进的特征描述方法，旨在增强对噪声的抵抗能力： 1. **相对顺序直方图（Histogram of Relative Orders in the Patch）**：这种方法考虑的是像素之间的相对排列顺序，而非绝对强度值。它能捕捉到图像中的结构信息，但在处理高斯噪声时，单纯依赖相对顺序可能不足以抵抗噪声的影响。 2. **局部二值模式（Local Binary Pattern, LBP）直方图**：LBP通过编码每个像素与其周围邻域的对比情况，形成二进制码，再统计这些码的分布。虽然LBP方法对光照变化有较好的不变性，但在高斯噪声下，原始的LBP编码也可能受到干扰。为了提高在高斯噪声环境下的鲁棒性，本文作者针对这两种方法进行了优化。他们着重于减少噪声对特征描述的影响，可能的改进策略包括： - **噪声抑制技术**：通过滤波器或降噪算法预处理图像，降低高斯噪声的干扰。 - **特征选择和强化**：挑选那些在噪声环境中依然稳定、具有高区分度的像素排列顺序或LBP码，作为描述子的一部分。 - **稳健的统计方法**：使用更稳健的统计模型来估计和量化直方图中的噪声成分，以提高描述符的噪声抗性。 - **匹配策略**：结合最近发展的高级技术，如局部特征匹配的细化算法，以提高不同描述符之间的匹配精度和稳定性。此外，论文还展示了如何将这些改进后的特征描述器应用于实际匹配任务中，以提升整体系统的性能和可靠性。这项工作为在高斯噪声背景下构建更强大的特征描述算法提供了新的思路和实用工具，对于提高计算机视觉系统在复杂环境中的表现具有重要意义。

Robust Order-based Methods for Feature Description

Raj Gupta, Harshal Patil and Anurag Mittal

Department of Computer Science and Engg.

Indian Institute of Technology Madras, Chennai INDIA - 600036

{rgupta,harshal,amittal}@cse.iitm.ac.in

Abstract

Feature-based methods have found increasing use in

many applications such as object recognition , 3D rec on-

struction and mosaicing. In this paper, we focus o n the

problem of matching such features. While a h istogram-of-

gradients type methods such as SI FT, GLOH and Shape

Context are currently p opular, several papers have sug-

gested using orders of pixels rather than raw intensities an d

shown improved results for some applications. The papers

suggest two different techniques for doing so: (1) A His-

togram of Relative Orders in the Patch and (2) A Histogram

of LB P codes. While these methods have shown g ood per-

formance, they neglect the fact that the orders can be quite

noisy in the presence of Gaussian noise. In this paper, we

propose changes to these approaches to make them robust

to Gaussian noise. We also show how the descriptors can

be matched using recently developed more advanced tech-

niques to obtain better matching performance. Finally, we

show that the two method s have complimentary strengths

and tha t by combining the two descriptors, one obtains

much better results tha n either of them con sid ered sepa-

rately. The re sults are shown on the standard 2D Oxford

and the 3D Caltech datasets.

1. Introduction

The use of features for image repr esentation and

matching has gained tremendous importance and popu-

larity in recent years for problems as diverse as Im-

age Alignment, mosiacing, 3D Reconstruction, Object

Recognition and Tracking. Features are extracted using

methods such as Harris features[4], Harris-afﬁne, Hes-

sian, Hessian-afﬁne[13] , M SER (Maximally Stable Ex-

tremal Regions)[20], DOG (Difference of Gaussians)[10]

and others[6, 25], metho ds for matching (normalized)

patches include the popular SIFT (Scale Invariant Fea-

ture Transform)[10] and its variants such as GLOH ( G ra-

dient Location and Orientation Histog ram)[15], Shape

Context[18] and other modiﬁcations of such gradient/edge

based methods such as [1, 7, 8, 18, 12].

More recently, methods have been proposed[11, 3] that

use orders of pixels rather than raw intensities and the re-

sults from these methods are encouraging. In this paper, we

propose new ways of using the orde r be tween pixels that are

more robust to noise in the underlying data. The two meth-

ods propo sed capture complementary properties of a feature

region - one captures the overall distribution of pixels in the

patch and the other captures local gradient properties.

The ﬁrst method is based on orders of pixels relative to

the entire patch and builds a histogram based on the rela-

tive position of the intensities w.r.t. to the entir e patch. The

second method looks at local orders of pixels and general-

izes the Center-Symmetric Local Binary Patterns (CS-LBP)

descriptor. Speciﬁcally, instead of a binary code, we de-

velop a ternary code, which we call Center Symm e tric Lo-

cal Ternary Patterns (CS-LTP). Both these methods are de-

signed to b e more robust to Gaussian noise than previously

considered descriptors. They captu re complementary infor-

mation and a combination of th ese two methods was found

to improve upon either of the two considered separately.

2. Related Work

The idea of matching images/features using order of in-

tensities rather than raw intensities is not new. By consid-

ering only th e orders be twe en pixels rather than their in-

tensities, one o btains invariance to a monotonic change in

the intensities. The Census algorithm[27] transforms the

intensity space to an “order” space, where a bit pattern is

formed by looking at the orders of a given pixel with its

neighbors. This algorithm essentially counts the numbe r of

ﬂipped point pairs in the patch. Bhat and Nayar[2] use an

improved version of this algorithm where they somewhat

alleviate the problem of counting even one salt-and-pepper

error in a pixel multiple times. Mittal and Ramesh [16] pro-

posed a method in which the penalty for a n order ﬂip is pro-

portion al to the intensity difference between the two ﬂipped

pixels. This reduces the error due to pixels whose order

may have got ﬂipped due to Ga ussian noise. Finally, Xie et.

al. [26] and Sing h et al [23] present a statistical approach

下载后可阅读完整内容，剩余7页未读，立即下载

emma9423

粉丝: 0
资源: 2

鲁棒像素顺序特征描述新方法对抗高斯噪声

《不稳定日志数据的鲁棒性缺陷检测模型研究》

实时目标跟踪算法优化：Robust scale-adaptive mean-shift

robust-redis-lock: 实现安全分布式锁的Ruby库

Feature Engineering for Time Series Forecasting: Experts Guide You in Building Forecasting Gold ...

MATLAB Control Systems Simulation: Detailed Steps for Building Models and Testing

Time Series Chaos Theory: Expert Insights and Applications for Predicting Complex Dynamics

MATLAB Genetic Algorithm: A Deep Dive into Bio-inspired Heuristic Optimization Techniques

深入解析Java版jesque-robust-worker-pool工人池实现机制

毕业设计基于单片机的室内有害气体检测系统源码+论文（高分毕设）

mellitz_3df_elec_01_220502.pdf

最新资源