Classificação de espécies de plantas a partir de imagens das componentes folha e flor.

Abstract

The classification of plant species is very challenging due to the biodiversity of our planet and our multiple ecosystems exponents this problem, for this, solutions over the years have been developed so that we can automate the work of plant classification, which before it was exclusive for the technical staff of the area. Convolutional Neural Networks and Deep Learning have been increasing the possibility of most suitable solutions for the task of classification making possible autonomous feature extraction. For this research, an image database belonging to the 2015 World Plant Recognition Challenge, LifeCLEF, this database has 113205 images of one thousand different species. In this work the classification of plant species was addressed using two components of the same plant, the leaf and the flower. To this end, pre-trained deep neural networks in the Imagenet database were used to classify each component, then different classifier combination rules were evaluated, creating a meta classifier responsible for merging. Experimental results allowed an increase of up to 23 percentage points (from 68% to 91 %) when the fusion of the leaf and flower components classifiers was performed.