Exploração de estratégias para a classificação de fluxos de dados de imagens ; Exploring strategies for image data stream classification

The image data stream classification presents several challenges, for example, the evolution of concepts of known classes (concept-drift) and the emergence of new classes (concept-evolution). Although many studies deal with the image data stream classification, these studies did not explore some characteristics of this context together. For example, specific evaluation methods for data stream scenarios, the evolution of the image descriptor (feature-evolution), updating the decision model considering characteristics of real application environments, and classification algorithms capable of dealing with high dimensional data. The work described herein aims to contribute to the image data stream classification exploring the stages of classification, model update, and evaluation, considering inherent aspects of real application scenarios. Therefore, the EVISClass framework was developed for the evaluation of algorithms for image data stream classification. This framework can consider: the occurrence of concept-drift and concept-evolution, delays for labeling images (latency), and active learning strategies for selecting instances to be labeled. The use of this framework allowed us to observe that latency has a strong influence on the efficacy of the results. Furthermore, we observed that active learning strategies could contribute to the selection of a smaller number of labeled instances without significantly impacting the classifier's effectiveness. Finally, the HubISC algorithm for the image data stream classification was developed. This algorithm incorporates the hubness aspect, which is inherent in high-dimensional data. The HubISC algorithm also provides a structure for summarizing instances using hubs, which are representative data instances. Furthermore, these instances are used in the algorithm as an active learning strategy. The experiment results with the HubISC algorithm show the potential in terms of predictive performance and the number of labeled instances compared to commonly used algorithms for ...