Recuperação de informação visual em bases de imagens de cidades históricas: contribuições para o reconhecimento e classificação de imagens

Abstract

This work tackles visual information retrieval for image datasets, regarding both scene recognition and image classification. Scene recognition is the task of recognizing a query image inside the dataset, matching their visual content. Concerning image classification, the goal is to separate dataset images into known categories. Those aspects of visual information retrieval are directly related to the organization of huge datasets and we improve the state-of-the-art for both, concerning specific applications, either performing descriptors filtering before image matching or using semantic regions for codifying images by visual dictionaries, respectively for image recognition and classification problems. Regarding scene recognition, our contribution is a methodology of enhancing the image matching algorithm through the use of subspace clustering algorithms. We present thus the aggregation of matching and clustering algorithms and, also devise a modified version of a literature subspace clustering, reducing its runtime while preserving the clusters discovery confidence. For the image classification issue, we develop a novel method which is based on both image codification by visual dictionaries and semantic regions. The proposed technique outperforms the state-of-the-art in all experiments.We employ such methods to evaluate literature image datasets and also a new dataset whose creation is explained in details, including image gathering, their selection and annotation. Scene recognition application follows the usual protocol of recognizing a dataset scene from a target image, consisting of an urban scene facade. For imageclassification we aim to classify architectural styles lying in a baroque city, separating baroque buildings from the contemporary ones.