局部差分二进制：超快速显著特征提取

需积分: 10 130 浏览量更新于2024-09-10 收藏 674KB PDF 举报

"这篇论文提出了一种新的特征描述符——局部差分二进制（Local Difference Binary, LDB），旨在解决计算机视觉应用中特征描述符的效率和独特性问题。LDB设计为一种快速且具有高区分度的二进制描述符，能够实现实时性能并能从大量经过各种变换的数据中准确识别匹配。通过简单的强度和梯度差异测试，LDB在图像块内的网格单元对上直接计算二进制字符串。此外，它还采用多网格策略和显著位选择方法来捕获不同空间粒度下的图像块独特模式。实验结果显示，与主要针对速度优化的现有最先进的二进制描述符相比，LDB在保持相似构建效率的同时，实现了更高的准确性并更快的速度。" 《局部差分二进制：超快显著特征提取》在计算机视觉领域，特征描述符的质量和效率是关键因素，直接影响用户的应用体验。现有的特征描述符通常存在两方面的问题：一是计算过程过于复杂，难以实现实时处理；二是独特性不足，导致在大规模数据库中进行匹配时容易出现错误。针对这些问题，该研究提出了局部差分二进制（LDB）描述符，它在保证高效的同时，提高了特征的识别精度。 LDB的核心思想是在图像块内采用简单强度和梯度差异测试，对成对的网格单元进行比较，生成二进制串。这种操作方式极大地简化了计算流程，使得LDB能够在短时间内完成特征提取。同时，为了增强特征的独特性，研究者引入了多网格策略，这允许在不同空间分辨率下捕获图像的特征，从而增加了描述符的表达能力。此外，LDB还采用了一种显著位选择方法，这是一种优化技术，旨在挑选出最具区分性的位，进一步提升匹配的准确性。这种方法确保了LDB在保持构建效率的同时，能够更有效地抵抗各种几何变换和光照变化。实验结果表明，与目前专注于速度优化的主流二进制描述符（如BRIEF、ORB等）相比，LDB在保持类似构建速度的基础上，其匹配准确性得到了显著提升，同时在运行速度上也有优势。这意味着LDB更适合于需要实时处理和高精度识别的计算机视觉任务，例如图像检索、目标检测和跟踪等。局部差分二进制（LDB）为特征描述符的设计提供了一个新的视角，它通过结合高效计算和高区分度，解决了传统描述符面临的挑战，有望成为未来计算机视觉领域的有力工具。

Local Difference Binary for Ultrafast

and Distinctive Feature Description

Xin Yang, Member, IEEE,and

Kwang-Ting (Tim) Cheng, Fellow, IEEE

Abstract—The efficiency and quality of a feature descriptor are critical to the user

experience of many computer vision applications. However, the existing

descriptors are either too com putationally expensive to achieve real-time

performance, or not sufficiently distinctive to identify correct matches from a large

database with various transformations. In this paper, we propose a highly efficient

and distinctive binary desc riptor, called local difference binary (LDB). LDB directly

computes a binary string for an image patch using simple intensity and gradient

difference tests on pairwise grid cells within the patch. A multiple-gridding strategy

and a salient bit-selection method are applied to capture the distinct patterns of the

patch at different spatial granularities. Experimental results demonstrate that

compared to the existing state-of-the-art binary descriptors, primarily designed for

speed, LDB has similar construction efficiency, while achieving a greater accuracy

and faster speed for mobile object recognition and tracking tasks.

Index Terms—Binary feature descriptor, mobile devices, object recognition,

tracking, augmented reality

1INTRODUCTION

FEATURE point descriptors are widely used in many computer

vision tasks such as marker-less augmented reality (AR) [10], [11],

simultaneous localization and mapping (SLAM) [16], and image

retrieval [19], [21]. Their broad applications have driven the

development of a plethora of descriptors [1], [2], [3], [4], [5], [6], [7],

[8]. However, as the application requirements are increasingly

demanding, for example, handling larger databases and/or

running real time on handheld devices, the demand for more

advanced descriptors is stronger than ever.

An ideal descriptor should achieve two competing goals: high-

quality description and low computational complexity. High-

quality descriptors capture the most representative information in

an image, such that different image content can be distinguished

(i.e., high distinctiveness) and the same content subject to various

image distortions can be recognized (i.e., high robustness). High-

speed descriptors enable the entire application task to run in real

time at a sufficiently high frame rate.

Many research efforts have been made to achieve either strict

quality requirements or low computational speed. The SIFT

descriptor [2], proposed over a decade ago, has been widely

adopted as one of the highest quality options. However, it imposes

a heavy computation burden. This drawback has drawn extensive

efforts [1], [3] for optimizing its speed without compromising its

quality too much. Among the enhancements, SURF [1] is arguably

the most noticeable. But recent experiments [22] have shown that

the SURF descriptor is still too computationally heavy; thus only a

limited number of points can be handled for real-time applications

such as AR, especially for handheld devices such as smartphones

and tablets. On the other end of the spectrum aiming primarily at

fast runtime, lightweight binary descriptors such as BRISK [6],

FREAK [7], BRIEF [4], and its variant rBRIEF (or ORB descriptor)

[5] have become increasingly popular as they are very efficient to

store and to match (simply computing the Hamming distance

between descriptors via XOR and bit-count operations). These

runtimeadvantagesmakethemmoresuitableforreal-time

applications and handheld devices. However, these binary

descriptors utilize overly simplified information, i.e., raw inten-

sities of a subset of pixels within an image patch for binary tests,

and thus have low discriminative ability. Lack of distinctiveness

incurs an enormous number of false matches when matching

against a large database. Expensive postverification methods (e.g.,

RANSAC [23] or PROSAC [13]) are usually required to discover

and validate matching consensus, increasing the runtime of the

entire process.

In this paper, we introduce a new binary descriptor, named

local difference binary (LDB), which achieves similar computa-

tional speed and robustness as the state-of-the-art binary descrip-

tors [4], [5], [6], [7], yet offering much higher distinctiveness

compared to them. The high quality of LDB is achieved through

three schemes. First, LDB utilizes both average intensity I

avg

and

first-order gradients, d

and d

, of grid cells within an image patch.

Specifically, the internal patterns of the image patch are captured

through a set of binary tests, each of which compares the I

avg

, d

and d

of a pair of grid cells (see Figs. 1a and 1b). The average

intensity and gradients provide a more complete description than

other binary descriptors. Second, LDB employs a multiple-

gridding strategy to encode the structure at different spatial

granularities (see Fig. 1c). Coarse-level grids can cancel out high-

frequency noise, while fine-level grids can capture detailed local

patterns, thus enhancing distinctiveness. Third, LDB leverages a

modified AdaBoost method to select a set of salient bits. The

modified AdaBoost targets the fundamental goal of ideal binary

descriptors: minimizing distances between matches while max-

imizing them between mismatches, optimizing the performance of

LDB for a given descriptor length. Computing LDB is extremely

fast. Relying on integral images, the average intensity and first-

order gradients of each grid cell can be obtained by only 4-8 add/

subtract operations.

Our experimental results demonstrate that the construction

speed of LDB is much faster than that of SURF and is comparable

to those of the state-of-the-art binary descriptors, including ORB,

BRISK, and FREAK, while the robustness and distinctiveness of

LDB is higher than these descriptors.

The remainder of this paper is organized as follows: Section 2

reviews the related work. Section 3 presents details of the

proposed descriptor. In Sections 4 and 5 we compare performance

of LDB with the state-of-the-art descriptors on public benchmarks

and evaluate its speed, robustness, and discriminative power for

mobile applications. Section 6 concludes the paper.

2RELATED WORK

SIFT is currently among the best quality descriptors in the

literature. It relies on local gradient histograms and represents an

image patch using a 128D real-value vector. Despite its high

descriptive power and robustness to a variety of image transfor-

mations, the intensive computations for obtaining gradients and

the high dimensionality of SIFT make it prohibitively slow to

compute and match, especially on low-power devices. PCA-SIFT

[3] reduced the descriptor from 128D to 36D to reduce the

matching cost, whereas the increased time for descriptor formation

almost annihilates the increased speed of matching. To date, SURF

descriptor is considered as the most popular replacement for SIFT.

It greatly accelerates the gradient computations using integral

images as [14], while almost preserving the quality of SIFT.

188 IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 36, NO. 1, JANUARY 2014

. The authors are with the Electrical and Computer Engineering Depart-

ment, University of California, Santa Barbara, Harold Frank Hall, Rm

4109, Santa Barbara, CA 93106-9560.

E-mail: xinyang@umail.ucsb.edu, timcheng@ece.ucsb.edu.

Manuscript received 9 Sept. 2012; revised 7 May 2013; accepted 21 July 2013;

published online 13 Aug. 2013.

Recommended for acceptance by T. Tuytelaars.

For information on obtaining reprints of this article, please send e-mail to:

tpami@computer.org, and reference IEEECS Log Number

TPAMI-2012-09-0713.

Digital Object Identifier no. 10.1109/TPAMI.2013.150.

0162-8828/14/$31.00 ß 2014 IEEE Published by the IEEE Computer Society

下载后可阅读完整内容，剩余6页未读，立即下载

lillllllll

粉丝: 69
资源: 13

局部差分二进制：超快速显著特征提取

ORB-快速特征点提取和描述的算法

LDB特征提取算法

UltraFast 设计方法

A Binary Firefly Algorithm-based Feature Selection:A Binary Firefly Algorithm-based Feature Selection-matlab开发

Salient local binary pattern for ground-based cloud classification

LBP-Learning-Multi-scale-Block-Local-Binary-Patterns-for-Face-Recognition.pdf

Binary Particle Swarm Optimization for Feature Selection：简单算法展示了如何在特征选择问题中使用二进制粒子群优化（BPSO）。-matlab开发

A Completed Modeling of Local Binary Pattern Operator for Texture Classification

Local Binary Features

local binary pattern

最新资源