Ensemble Classifier Design applied to face expression classification

模式分类涉及将对象分配到几个预先指定的类别或类之一，是许多数据解释活动中的关键组成部分。所提出的方法侧重于从示例中学习的分类器，并假设每个示例模式由一组数字表示，这些数字称为模式特征。在面部表情分类的情况下（例如区分微笑和皱眉的脸），这些特征可以由代表面部特征不同方面的数字组成。为了设计系统，通常将示例模式分为两组，一个是设计分类器的训练集，另一个是随后用于预测应用以前未见过的示例时的性能的测试集。当有很多特征和相对较少的例子时，问题就出现了，分类器可以很好地学习训练集，称为过度拟合，因此测试集的性能下降。集成分类器领域的发展是为了解决使用相对简单的分类器组合来实现最佳模式分类性能的问题。人们已经发现,组合的优点在于,它不太可能over-fit。然而，调整单个分类器仍然存在困难，这个过程通常是使用分类器参数执行的（例如神经网络分类器的复杂性）。常见的方法是进一步划分训练集以产生可用于调整适当参数的验证集。然而，当示例数量不足时，这些技术要么不合适，要么非常耗时。在最近的工作中，首席研究员开发了一种集成类可分离性度量，该度量在训练集上计算，可以检测过拟合。因此，不需要验证集，从而使更多的数据可用于训练。项目建议是在实际数据上对方法进行测试，并在基准数据上确认之前得到的结果。提出了两类问题的纠错输出编码方法，并提出了多类问题的纠错输出编码方法。ECOC是一种将多类问题分解为两类子问题的集成技术。在本提案中，目的是理解为什么该技术工作良好，并提出一种设计方法，目的是将其应用于面部表情分类问题。进一步的目标是将该方法应用于仅使用训练集来预测特征选择中的最优特征数量。多年来，人们一直在努力发现最相关的特征，因为结果已被证明是更准确和有效的分类器。一些早期的研究表明，使用类可分性度量是一种可行的方法。当有数百或数千个特征时，这个问题尤其具有挑战性，就像在某些生物识别、生物信息学和数据挖掘应用中一样。众所周知，即使是模式分类系统性能的微小改进也会影响商业可行性，项目的成功结果也会影响其他生物识别、生物信息学和数据挖掘应用。拟议的研究与EPSRC的任务相关，因为它旨在推进具有实际应用相关性的知识和技术。预计这项研究的可能结果将是通过利用这项技术提高英国的竞争力。在项目合作伙伴三菱电机的帮助下，开发的技术将应用于物理安全系统的应力分析和汽车应用的驾驶员疲劳。

Pattern classification involves assignment of an object to one of several pre-specified categories or classes, and is a key component in many data interpretation activities. The proposed approach focuses on classifiers that learn from examples, and it is assumed that each example pattern is represented by a set of numbers, which are known as the pattern features. In the case of face expression classification (for example distinguish between a smiling and frowning face), these features could consist of numbers representing different aspects of facial features. In order to design a system it is customary to divide the example patterns into two sets, a training set to design the classifier and a test set which is subsequently used to predict the performance when previously unseen examples are applied. A problem arises when there are many features and relatively few examples, and the classifier can learn the training set too well, known as over-fitting so that performance on the test set decreases. The field of ensemble classifiers has been developed to address the problem of achieving the best pattern classification performance using a combination of relatively simple classifiers. It has been found that the combination has the advantage that it is less likely to over-fit. However, there is still the difficulty of tuning the individual classifiers, a process that is normally performed using classifier parameters (for example complexity of a neural network classifier). The common approach is to further divide the training set to produce a validation set that can be used to adjust appropriate parameters. However, when the number of examples is in short supply theses techniques are either inappropriate or very time-consuming.In recent work, the Principal Investigator has developed an ensemble class separability measure that is computed on the training set and that can detect over-fitting. Therefore there is no need for a validation set, thereby making more data available for training. The project proposal is to test the method on real data and confirm the results that have been obtained previously on benchmark data.The technique was proposed for two-class problems, and the proposal is to develop the method for multi-class problems using Error-Correcting-Output-Coding (ECOC). ECOC is an ensemble technique that works by decomposing a multiclass problem into two-class sub-problems. In this proposal the aim is to understand why the technique works well, and to propose a design methodology with the aim of applying it to problems in face expression classification. A further objective is to apply the method to predicting the optimal number of features in feature selection using only the training set. There has been for many years a great deal of effort in discovering the most relevant features, since the result has been shown to be more accurate and efficient classifers. Some early research indicates that using class separability measure, this is a feasible approach. The problem is particulary challenging when there are hudreds or thousands of features,as there are in certain biometric, bio-informatics and data mining applications.It is known that even a small improvement in performance of a pattern classification system can affect commercial viability, and the successful outcome of the project should impact other biometric, bio-informatics and data mining applications. The proposed research is relevant to the EPSRC mission since it is aimed at advancing knowledge and technology with practical application relevance. It is anticipated that the likely result of this research will be the enhancement of UK competitiveness through exploitation of the technology.With the help of project partner Mitsubishi Electric, the developed techniques will be applied to stress analysis for physical security systems and driver fatigue for automotive applications.