Combinação estática e dinâmica de classificadores na tarefa de classificação de texturas

Abstract

Texture classification is a task with applications across multiple domains, ranging from facial recognition to cancer detection in medical imaging, among others. In traditional approaches, the success of the application relies heavily on the feature extraction and classification steps. Over the years, numerous feature extraction methods have been proposed, with non-handcrafted (deep learning-based) approaches consistently outperforming handcrafted (manually designed) ones. This work investigates static and dynamic selection techniques of classifiers trained with features extracted from non-handcrafted architectures. For the experiments, two widely recognized texture classification datasets were used: the Flickr Material Database (FMD) and the Describable Texture Dataset (DTD). Support Vector Machines (SVMs) with grid search for hyperparameter tuning and 10-fold cross-validation were employed. We initially evaluated individual features and found that Visual Transformers (ViT) outperformed other architectures. Performance improvements were also observed by combining top-performing individual classifiers using static combination methods, such as sum, product, and maximum rules. On the other hand, dynamic classifier selection techniques did not yield the expected improvements. The most significant results were achieved using static classifier combinations: on the FMD dataset, we achieved an F1-score of 93.2% by combining ViT models using the product rule, while for the DTD, we reached 81.7% by using the top five individual classifiers with the sum rule. The advances obtained through this work position our approach among the top five state-of-the-art methods for both datasets.