Abstract
In this paper, we proposed an application of semi-supervised image classification method with Boolean type supervised data that collected via smartphone. The image classification is a difficult problem for the machine learning methods because there is no effective feature extraction method for practical use. Thus, the supervised classification approaches are necessary to classify images. However, it is necessary to large cost for the label instruction when the number of data samples is very huge. To solve this problem, we focus on the semi-supervised approaches. The semi-supervised learning is a learning method using the data which include labeled samples and unlabeled samples. The accuracy of semi-supervised learning depends on the number of labeled samples. Therefore we proposed an interface application for semi-supervised image classification via smartphone. Recently, the touch panel interface is commonly used for input device of smartphone. Considering this, we used simple supervised information, such as "Similar / Dissimilar", to the similarity between different 2 images. The images are classified by hierarchical clustering with a distance function that derived by supervised information. However, it is difficult to completely make the supervised samples in huge data. To be able to calculate distance, dissimilarity of non-supervised samples is estimated from the machinery extracted features such as brightness, hue, shape and etc. In this experiment, we create an image classification problem data set adapted to the practical use and investigate the effectiveness of proposed image classification method on our data set.
