超像素生成算法的噪声鲁棒性研究

论文总字数：24313字

摘 要

关键词……………………………………………………………………………………………1

Abstract……………………………………………………………………………………………2

Key words…………………………………………………………………………………………2

1 绪论……………………………………………………………………………………………3

1.1 计算机视觉学科简介………………………………………………………………………3

1.2 超像素概念的提出…………………………………………………………………………3

1.3 超像素生成算法的类别及各自的主要性质………………………………………………4

1.4 超像素分割的意义及实际应用……………………………………………………………6

1.5 本文主要研究内容…………………………………………………………………………6

1.6 论文组织框架………………………………………………………………………………7

2 相关概念简介…………………………………………………………………………………8

2.1 用于分析的3种超像素生成算法……………………………………………………………8

2.1.1 SLIC算法…………………………………………………………………………………8

2.1.2 TurboPixels算法……………………………………………………………………………9

2.1.3 基于熵率的方法（ER）…………………………………………………………………11

2.2 用于评价超像素生成算法的定量分析指标………………………………………………13

2.2.1 边缘召回率（Boundary Recall）…………………………………………………………13

2.2.2 欠分割错误率（Under-Segmentation Error）……………………………………………14

2.2.3 可完成的分割精度（Achievable Segmentation Accuracy）……………………………16

3 本文工作内容…………………………………………………………………………………17

3.1 实验材料……………………………………………………………………………………17

3.2 实验方法……………………………………………………………………………………17

3.3 实验数据……………………………………………………………………………………18

3.4 实验结论……………………………………………………………………………………25

4 基于MATLAB GUI的超像素生成算法噪声鲁棒性研究系统……………………………27

4.1 图形用户界面GUI简介……………………………………………………………………27

4.2 系统运行环境………………………………………………………………………………28

4.3 系统功能演示………………………………………………………………………………28

4.3.1 使用不同超像素生成算法分割单幅图片………………………………………………29

4.3.2 按目录批量生成定量分析所用的标号（label）文件……………………………………31

4.3.3 超像素生成算法的定量分析指标计算…………………………………………………32

5 总结……………………………………………………………………………………………34

5.1 研究成果总结………………………………………………………………………………34

5.2 可改进的方面………………………………………………………………………………35

致谢………………………………………………………………………………………………36

参考文献…………………………………………………………………………………………37

超像素生成算法的噪声鲁棒性研究

学号：09012232 学生姓名：郑颖

指导教师：孔佑勇

摘要：“超像素”这一概念最早于2003年由Xiaofeng Ren在一篇关于计算机视觉领域研究的论文中被提出。它可以被描述为具有相似颜色、亮度、纹理等特征的相邻像素所构成的图像块。超像素可以大幅度地提升计算效率，实际应用当中的图像通常会存在一些具有极大相似性的像素，如果在分析和处理这些像素时将它们视为一个整体，将会大幅度地降低计算复杂度。因此，研究超像素生成算法以及不同生成算法的性质在计算机视觉领域中具有相当重要的意义，而噪声鲁棒性是超像素生成算法的重要性质之一。本文采取将原始图像与噪声图像相合成的方法，对比不同超像素生成算法对于原始图像与噪声合成图像的定量分析指标差异，从而对不同超像素生成算法的噪声鲁棒性作出结论。

关键词：计算机视觉；超像素生成算法；定量分析；噪声鲁棒性

A research about noise robustness of superpixel algorithms

Student name: ZHENG Ying

Supervisor: KONG Youyong

Abstract:

The concept of "superpixel" was firstly purposed in a paper about a research of computer vision by Xiaofeng Ren in 2003. It can be described as a kind of image blocks composed of neighbouring pixels whose features such as colours, brightnesses and textures are similar. The idea of "superpixel" can improve the computational efficiency significantly. Images in practical applications generally consist of pixels which are tremendously similar. If they are regarded as a whole when being analysed and processed, the computation complexity will be significantly reduced. Thus, researches about superpixel algorithms and their different properties are of paramount significance to computer vision. Robustness is one of the important properties of superpixel algorithms. In this paper, we will compose the original image and the noise image to obtain the noise-composed image, calculate the difference value between quantitative analysis indexes respectively calculated by original image and noise-composed image of a superpixel algorithm, and then compare the difference values of different superpixel algorithms, therefore draw a conclusion about the noise robustnesses of different superpixel algorithms.

Key words: computer vision, superpixel algorithm, quantitative analysis, noise robustness

剩余内容已隐藏，请支付后下载全文，论文总字数：24313字