Multi-Task Ensemble Creation for Advancing Performance of Image Segmentation

Image classification is a special type of applied machine learning tasks, where each image can be treated as an instance if there is only one target object that belongs to a specific class and needs to be recognized from an image. In the case of recognizing multiple target objects from an image, the image classification task can be formulated as image segmentation, leading to multiple instances being extracted from an image. In the setting of machine learning, each instance newly extracted from an image belongs to a specific class (a special type of target objects to be recognized) and presents specific features. In this context, in order to achieve effective recognition of each target object, it is crucial to undertake effective selection of features relevant to each specific class and appropriate setting of the training of classifiers on the selected features. In this paper, a multi-task approach of ensemble creation is proposed. The proposed approach is designed to first adopt multiple methods of multi-task feature selection for obtaining multiple groups of feature subsets (i.e., multiple subsets of features selected for each class), then to employ the C4.5 algorithm or the KNN algorithm to create an ensemble of classifiers using each group of feature subsets resulting from a specific one of the multi-task feature selection methods, and finally all the ensembles are fused to classify each instance. We compare the performance obtained using our proposed way of ensemble creation with the one obtained using classifiers trained on different feature sets prepared through various ways. The experimental results show some advances achieved in the overall classification performance through using our proposed ensemble creation approach, in comparison with the use of existing feature selection methods and learning algorithms.