Quantitative descriptors for a range of visual biologic pigmentation patterns

• Gabriel Henrique Moro UFRGS
• Marcelo de Gomensoro Malheiros FURG
• Marcelo Walter UFRGS

Resumo

Computer Graphics has successfully modeled and rendered many natural patterns, such as the spots on mammals, veins on leaves, and many others. Nevertheless, a fundamental problem remains: how to validate the results beyond subjective visual similarity. We propose a small set of quantitative descriptors tailored to describe the most common pigmentation biological patterns: spots, labyrinths, and stripes, found in many species of mammals and fishes. Our goal is not to classify images blindly, which can be done through deep neural networks but to gain insights into the structure of animal skin pigmentation. For that, we pursue specialized descriptors in the same sense as the general ones from content-based image retrieval systems. First, we compute twelve metrics using image processing tools from a set of real patterns curated from scratch from publicly available repositories. We then normalize those measures by the average and calculate the variance to yield our descriptors. We validate the descriptor’s usefulness through two machine learning tasks on an augmented dataset with real and synthetic patterns. First, the descriptors are used as features of a supervised classifier with an overall accuracy of 98.4%. Second, we test the descriptors in an unsupervised clustering task, differentiating natural from artificial patterns and correctly identifying species present in the natural set. These tests show that the proposed small set of descriptors is representative of the pattern domain analyzed.

Palavras-chave: Deep learning, Visualization, Veins, Transfer learning, Pipelines, Neural networks, Skin