MULTI-PERSPECTIVE CREATION OF DIVERSITY FOR IMAGE CLASSIFICATION IN ENSEMBLE LEARNING CONTEXT

Image classification is a special type of classification tasks in the setting of supervised machine learning. In general, in order to achieve good performance of image classification, it is important to select high quality features for training classifiers. However, different instances of images would usually present very diverse features even if the instances belong to the same class. In other words, one types of features may better describe some instances, whereas other instances present more other types of features. The above description can indicate that the same learning algorithm may be capable of learning from some parts of a data set but show weaker ability to learn from other parts of a data set, given that different algorithms usually show different suitability for learning from instances that show various characteristics. On the other hand, image features are typically in the form of continuous attributes which can be handled by decision tree learning algorithms in various ways, leading to diverse classifiers being trained. In this paper, we investigate diversified adoption of the C4.5 and KNN algorithms from different perspectives, such as diversified use of instances and various ways of handling continuous attributes. In particular, we propose a multi-perspective approach of diversity creation for image classification in the setting of ensemble learning. We compare the proposed approach with those popular algorithms that are used to train classifiers on either a full set of original features or a subset of selected features for image classification. The experimental results show that the performance of image classification is encouraging through the adoption of our proposed approach of ensemble creation.