偏标记学习算法及其在图像分类中的应用研究毕业论文
2021-11-09 21:13:56
摘 要
随着科学技术的发展,人工智能技术正融入进人们生活的方方面面。机器学习是人工智能的重要分支,而弱监督学习是机器学习领域中一种十分贴近现实生活条件的场景。偏标记学习是弱监督学习领域里的一类重要框架,该框架下学习算法可获得介于完全监督学习和无监督学习之间的监督信息。
本文利用偏标记学习算法对图像分类进行研究,对偏标记学习框架下数据降维进行积极探索,主要工作如下:
(1)研究偏标记学习框架,对其中的三种典型算法PL-NN、PL-ECOC和IPAL算法进行了分析;
(2)分析机器学习中常见的维度灾难问题,在偏标记学习框架下基于DELIN算法引入核函数的概念,研究了一种新颖的降维算法——Kernel based Disambiguation Enabled LINear discriminant analysis(KDELIN);
(3)结合偏标记学习算法与KDELIN算法,将其应用于图像分类与人脸年龄估计中,在三种经典图像数据集上开展实验,验证了KDELIN是一种十分有效的降维学习策略,显著提高了学习器的泛化性能。
本文的贡献在于在数据降维中引入了核函数的思想,通过迭代进行核化的Fisher判别分析与候选标记消歧完成偏标记数据集的降维,研究工作表明偏标记学习算法能够有效地完成图像分类任务,并在与偏标记降维算法结合后显著提高了分类准确率。
关键词:偏标记学习;数据降维;核函数;Fisher判别分析;图像分类
Abstract
With the development of science and technology, artificial intelligence technology is being integrated into every aspect of people's life. Machine learning is an important branch of artificial intelligence, and weak supervised learning is a very close to the real life conditions in the field of machine learning. Partial label learning is an important framework in the field of weak supervised learning. Under this framework, the learning algorithm can obtain the supervised information between completely supervised learning and unsupervised learning.
In this paper, image classification is studied by means of partial label learning algorithm, and the dimensionality reduction problem under partial label learning framework is actively explored. The main work is as follows:
- The partial label learning framework is studied, and three typical algorithms PL-NN, PL-ECOC and IPAL are analyzed.
- Common dimension disaster problem in machine learning is analysed. In partial label learning framework, a novel Kernel-based Disambiguation Enabled LINear discriminant analysis (KDELIN) algorithm is proposed;
- Combined with the KDELIN algorithm, the partial label learning algorithms are applied to image classification and facial age estimation. Experiments are carried out on three typical image data sets. It is verified that KDELIN is a very effective dimension reduction learning strategy, which significantly improves the generalization performance of the learning algorithm.
Contribution of this paper mainly is in data dimension reduction in partial label learning framework. In the procedure of dimension reduction, KDENLIN is accomplished by alternating kernel-based Fisher discriminant analysis and candidate label disambiguation iteratively. The experiment results show that partial label learning algorithm can effectively accomplish image classification task, and partial label dimension reduction algorithms significantly improve the classification accuracy of learning algorithms.
Key Words:Partial Label Learning;Dimensionality Reduction;Kernel Function;Fisher Discriminant Analysis;Image Classification
目 录
第1章 绪论 1
1.1 引言 1
1.2 研究现状 3
1.3 研究内容 3
1.4 本文组织 4
第2章 偏标记学习算法 5
2.1 问题定义 5
2.2 算法描述 5
2.2.1 PL-NN算法 5
2.2.2 PL-ECOC算法 6
2.2.3 IPAL算法 8
2.3 本章小结 10
第3章 基于核函数的偏标记学习降维策略 11
3.1 背景描述 11
3.2 KDELIN降维策略 11
3.2.1 核化FDA降维 11
3.2.2 候选标记消歧 14
3.3 本章小结 16
第4章 实验及结果分析 17
4.1 数据集及评测指标 17
4.1.1 数据集介绍 17
4.1.2 评测指标 18
4.2 实验设计 18
4.3 实验结果 18
4.3.1 基于LOST数据集的对比实验 18
4.3.2 基于FG-NET数据集的对比实验 19
4.3.3 基于MIRFlickr数据集的对比实验 20
4.4 本章小结 22
第5章 总结与展望 23
5.1 工作总结 23
5.2 未来展望 23
致 谢 24
参考文献 25
附 录 27
绪论
引言
随着科学技术的发展,人工智能技术融入进生活的方方面面,其广泛的应用改变了社会的生产方式,改善了人们的生活质量。人工智能的未来正吸引着越来越多的科技企业和研究工作者的关注。在人工智能界,机器学习被认为是人工智能领域中最能够体现智能的一个分支。从历史来看,机器学习也是发展最快的一个分支。
机器学习是这样一门学科,它通过计算的手段,利用经验来改善系统自身的性能[1]。Mitchel对机器学习给出了更加形式化的定义[2] :假设性能评估手段P用来评测某段程序在任务T上的性能,如果计算机程序通过在经验E上的学习,其性能得到了提升,那么称该程序在E上进行了学习。机器学习的两大问题包括分类与回归。根据监督信息的不同,机器学习主要包括监督学习(Supervised Learning),无监督学习(Unsupervised Learning)以及弱监督学习(Weakly Supervised Learning)。