Técnica aprimorada de segmentação não-supervisionada em imagens com felinos domésticos

Abstract

A great number of recent projects have, as their focus, the preservation of fauna and flora through monitoring and research centered on regions with very heterogenic ecosystems, such as the Amazon Rain Forest. In fact, research projects based on animal monitoring are carried out in various parts of the world. The main problem of this type of monitoring lies on the cataloguing aspect that is still completed manually, consuming precious time of the researchers which could be better used in truly achieving the objectives of the research. As an example, in Australia, the lack of monitoring in several species of felines, especially domestic cats, is a concern of scientists because of questionable decisions made by governments that consider these animals as pests and treat them as a menace for environment balance. In Brazil, similar researches are conducted in order to maintain conservation of wild cats, such as jaguars. In this context, the objective of this work is to collaborate in this area with the study of pattern recognition and digital image processing in order to build a more effective method for animal segmentation in pictures, particularly the domestic cat. This method consists in creating a combined process that integrates a contrast enhance Color Boost filter, homomorphic filter, Mean- Shift filter and Distance Map in order to achieve an unsupervised way for segmenting cats on picture scenes. In addition to this method, a merge rule for decreasing the process of over-segmentation in images is applied, avoiding this common issue in many Watershed algorithms. The results can reach up to 84% on average accuracy in feline segmentation, with the possibility, in the future, to be extrapolated to others objects or species.