Adaptive Learning of Classifiers and Its Applications

为了识别图像中的目标对象，经常利用从示例中学习的技术。为了学习好的分类器，通常需要大量的样本。另一方面，人类可以有效地识别对象，即使它是一个新的对象。为了在机器上实现这种能力，有必要开发一种学习方法，从非常小的学习示例中给出在识别率和识别速度方面都有效的分类器。基于非线性函数的分类器常被用来识别图像中的模式。虽然它们的分类性能相当好，但它们的计算成本往往令人望而却步。为了克服这个困难，在这份报告中，我们首先提出了一种方法来识别模式的图像的基础上的分层分类器。它由两类分类器组成：简单快速分类器，其性能不一定令人满意;非线性慢速分类器，其性能相当好。通过使用第一（底部）分类器，许多假目标被拒绝。第二，复杂的分类器被应用于其余的模式。通过这种架构，快速和可靠的分类器成为可能。作为非线性分类器，我们使用RBF-SVM。在SVM的情况下，识别速度取决于支持向量的数量。为了进一步提高识别速度，我们提出了一种基于支持向量采样的SVM近似方法。我们还提出了方法来学习快速和可靠的分类器的基础上的提升算法的RBF中心的数量的上限。为了从有限数量的例子中学习分类器，我们提出了一种利用虚拟例子的方法。在报告中，我们首先提出了基于简单几何变换的综合方法。然后，为了提高手写模式样本的质量，我们还提出了一种从在线数据中生成模式的方法。

To recognize target objects in the image, technique of learning from examples are often exploited. To learn the good classifier, usually a lot of samples are required. On the other hand, human beings can recognize an object efficiently even if it is a novel object. To realize such ability on machines, it is necessary to develop a learning method to give classifiers which are efficient with respect to both recognition rate and recognition speed, from very small learning examples. Classifiers based on the non-linear functions are often used to recognize pattern in a image. Although their classification performance is quite good, their computational cost is often prohibitive. To overcome the difficulty, in this report, we first propose a method to recognize patterns in a image based on the hierarchical classifiers. It consists of two kinds of classifiers : simple and fast classifier, whose performance is not necessarily satisfactory, and the non-linear slow classifier, whose performance is quite good. With the use of first (bottom) classifier, many false targets are rejected. The second, complex classifier is applied to the remaining patterns. By this architecture, fast and reliable classifiers are made possible. As the non-linear classifiers, we use RBF-SVM. In the case of SVM, recognition speed depends on the number of support vectors. To improve the recognition speed further, we propose a method to approximate the SVM based on the sampling from the support vectors. We also propose methods to learn fast and reliable classifiers based on the boosting algorithm subject to the upper bound of the number of RBF centers. To learn the classifiers from the limited number of examples, we propose a method to utilize virtual examples. In the report, we first propose the synthesis method based on the simple geometric transformation. Then, to improve the quality of the samples for hand written pattems, we also propose amethod to generate pattems from online data.