Identificação de espécies de plantas utilizando autoencoder convolucional e aprendizagem não supervisionada

Abstract

Plants play a fundamental role in the existence of life on planet Earth, as they convert carbon dioxide (CO2) into oxygen (O2) and serve as food for most living beings, in addition to being used by various industrial segments. The importance of works in the research line of identification/classification of plant species is due to the vast biodiversity, in which many of these are at risk of extinction or even have not been scientifically cataloged/discovered. Still, there is the difficulty of performing the classification tasks manually. Studies show that the automated form of classification has been efficient, its processes demand less time and amount of work for the researcher, thus obtaining good results in the classification and labeling of botanical species. In this paper, an artificial neural network known as auto-encoder was used, specifically the convolutional auto-encoder, which employs the unsupervised/selftaught learning method, using unlabeled databases, as these are easier to be found digitally, to perform the training of computational models with images from a different domain and belonging to the same domain. Afterwards, the trained models were used to generate representations of different characteristics of the Flavia, Leafsnap and PlantCLEF2015 bases, which were used to train classifiers of the SVM type, individually reaching hit rates of up to 95,00%. Combination methods of classifiers were also used, showing themselves capable of achieving results that are competitive with those presented in the state of the art.