基于迁移学习的表情识别算法研究

发布时间：2018-05-09 19:59

本文选题：迁移学习 + 域适应　；参考：《南京邮电大学》2017年硕士论文

【摘要】：随着大数据时代的到来,人们可以更加容易地获得大量数据。此外,由于机器学习领域不断的发展,如何让计算机具有举一反三的能力,如何使大量数据可以更好地发挥作用,这些问题均变得非常实际和有价值。为了解决这些问题,迁移学习被提出并越来越受到人们的重视。在常规机器学习中有一个重要假设,即训练所需的数据和目标所需的数据必须具有相同的分布或者来自相同的特征空间。在现实生活中,这一假设是很难实现的。具体来说,对于一个分类问题,如果训练集的样本和目标集的样本不具有相同的分布,这就可以理解为源域与目标域不具有相同的特征空间。传统的解决方法是通过搜集更多的数据去模拟目标域的分布,但是这样代价极大。迁移学习是一种有效缩短两个域之间“距离”的方法,这样就可以接近传统机器学习中的假设,通过源数据训练出适合目标数据的模型。论文对迁移学习算法做了详细的研究和总结。具体工作如下:(1)对现有迁移学习算法做了全面的总结和研究,并对各种方法的性质进行了比较,对各种算法适合使用的领域也做了详细的阐述。(2)具体研究了迁移成分分析算法、测地流核算法、子空间对齐算法、最大独立域适应算法、信息理论学习算法等常用的算法,同时还研究了基于深度学习的迁移学习方法。(3)将迁移成分分析算法、测地流核算法、子空间对齐算法、最大独立域适应算法、信息理论学习算法以及深度学习相关的迁移学习算法运用到人脸表情识别当中,并使用不同的数据库对迁移学习算法在人脸表情识别中做了实验和比较。迁移学习有效解决了人脸表情识别当中源域与目标域不具有相同特征空间时的分类问题。(4)对迁移学习中存在的问题进行了分析,并对未来的发展进行了展望。
[Abstract]:With the arrival of big data's era, people can easily access a large number of data. In addition, due to the continuous development in the field of machine learning, how to make the computer have the ability to draw inferences from one another and how to make a large number of data work better, these problems have become very practical and valuable. In order to solve these problems, transfer learning has been put forward and paid more and more attention to. There is an important assumption in conventional machine learning that the training data and the target data must have the same distribution or from the same feature space. In real life, this assumption is difficult to achieve. Specifically, for a classification problem, if the samples of the training set and the target set do not have the same distribution, it can be understood that the source domain and the target domain do not have the same feature space. The traditional solution is to simulate the distribution of target domain by collecting more data, but this is costly. Transfer learning is an effective method to shorten the "distance" between two domains, so that it can approach the hypothesis of traditional machine learning and train a model suitable for target data through source data. In this paper, the transfer learning algorithm is studied and summarized in detail. The specific work is as follows: (1) A comprehensive summary and study of the existing transfer learning algorithms are made, and the properties of the various methods are compared, and the fields in which the algorithms are suitable for use are also elaborated in detail. (2) the migration component analysis algorithms are studied in detail. Geodesic flow accounting method, subspace alignment algorithm, maximum independent domain adaptation algorithm, information theory learning algorithm and other commonly used algorithms. At the same time, the migration component analysis algorithm, geodesic flow accounting method, is also studied. Subspace alignment algorithm, maximum independent domain adaptation algorithm, information theory learning algorithm and transfer learning algorithm related to depth learning are applied to face expression recognition. Different databases are used to compare the transfer learning algorithm in facial expression recognition. Migration learning effectively solves the classification problem in facial expression recognition where the source domain and the target domain do not have the same feature space.
【学位授予单位】：南京邮电大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP181

【相似文献】