Explorando técnicas de aprendizado híbrido para o reconhecimento automático de imagens de plantas

Abstract

In recent years, with the evolution of Convolutional Neural Networks (RNC’s), the automatic recognition of plant species from images has become a very relevant research topic for scientists, researchers and students in both the botany and computer communities. The main challenges involving automatic recognition of plant species are directly related to intra-class variability and inter-class similarity, both arising from the complexity of the images in question. The main objective of this work is to explore hybrid learning techniques, that is, the combination of supervised and unsupervised learning techniques with the purpose of minimizing the impacts of this variability in the process of classification of plant images. In this work we explored the use of features extracted by RNC’s for the recognition of plant species images. As object of study, the ImageCLEF2013 (PlantCLEF) image collection with 26,077 images from 250 plant species was used. Partitioning techniques were applied to each base with different features. In addition, more comprehensive classification approaches were explored, using classical methods such as the Random Forest algorithm and eight variations of the SVM classifier with features extracted by RNC’s Inception V3, VGG-16 and VGG-19. Nevertheless, the Softmax classifier layer of each of the RNC’s was also considered in order to verify the impact of partitioning on the image recognition process of plant species. As a result, the experiments show it is possible to improve the results of classifier effectiveness by combining feature extraction by RNC’s and class partitioning with grouping techniques.