Identificação de golfinhos através de Técnicas de Processamento Inteligente de Imagens

Abstract

The inteligente processing of digital images applied to the identification of dolphins is based on several tecniques as preprocessing and segmentation of images, neural networks, mathematical morphology and geometrical transformation; they are combined to allow the dolphin images to be treated to identify different specimen. The analyzed dolphins in this work are from the S. fluviatilis species usually known as boto-cinza in Portuguese, and live in a sea area between Pipa and Tabatinga, two beaches in Rio Grande do Norte - Brazil. One of their main characteristics is a large dorsal fin. This fin acts as a dolphin identity card, since the aggressions suffered by the animal stay registered there. Another important characteristic is the own shape of the fin. In this work, we have implemented several algorithms to preprocess the image using the KLT theory and auto-contrast. We have used mathematical morphology tecniques to help extracting the characteristics and neural networks to classify the specimen. To identify the fin curvature we used two parametric cubic polynomials, associated to the two sides of the fin and both originating at the fin peak. Notches are detected at each curve and mapped to the parametric space of the polynomials. Some remarkable points are informed by the user to normalize the polynomials. The final data are maintained in a database and used to identify new images. Experimental results show that the selected characteristics are good identifiers, correctly classifying different specimens and recognizing with few errors the same specimen in different images.