Aprendizagem e recuperação de imagens utilizando mapas auto-organizáveis e representação log-polar.

Abstract

Everyday, a growing number of organizations is collecting and storing a large amount of digital images. Furthermore, images have also been massively added to the World Wide Web. Therefore, structuring this information, in order to allow efficient retrieval, is a very important task. In the initial image retrieval systems, an image indexing scheme based on keywords was used. However, due to the fast growth of the digital image collections, two problems became evident: (i) the vast amount of labor required in manual image annotation and (ii) the subjectivity of human perception. Thus, in the 1990’s, the idea of Content Based Image Retrieval (CBIR) emerged, which is characterized by automatically indexing images using their own visual features, such as color, texture and shape. Many indexing methods based on B-Trees have been used in CBIR, in order to reduce the search time. However, these methods are generally inefficient when working with high dimensions. Moreover, the feature extraction techniques can cause the loss of valuable image information. In this work, we investigate the use of Neural Networks (more specifically, the Self-Organizing Maps) to classify, index and retrieve image in this class of problems. The image representation (log-polar) adopted in this work helps recognizing objects in a way that is independent of orientation and scale, besides being more compact than the original input image. The experimental results (related to individual objects and arbitrary image recognition) showed that the combination of Self-Organizing Maps with the log-polar representation is a promising strategy for image classification. Thus, a CBIR system prototype was built using the proposed strategy and applied to two case studies of image retrieval from the Web.