Análise de arquiteturas de redes neurais siamesas para a classificação de espécies de plantas.

Abstract

The classification of plant species, specifically those that use leaf images, has been a great challenge, which requires the analysis of experts in botany. It may be related to a large number of plant species already cataloged and the fact that some plant leaves are similar, despite belonging to different species. To facilitate the plant classification process, automatic systems that use machine learning and computer vision techniques to differentiate species have been proposed, especially Convolutional Neural Network (CNN) models, which have been widely used for feature extraction and classification of plant species. However, the use of CNN requires a large number of images to carry out its training. Also, this tool is not scalable, so if a new class is added, the network needs to retrain. Thus, an alternative has been the use of Siamese Neural Networks (SNN). This study aims to evaluate different architectures of Siamese Neural Networks to plant species classification from leaf images. Additionally, the use of features extracted from the intermediate layers of SNNs is investigated, observing the impact on the results and seeking to prioritize approaches that are efficient even with few training images. Experiments on Flavia and MalayaKew leaf image databases have shown that the fusion of intermediate features improves SNN performance. For SNNs composed of VGG16, MobileNet, and DenseNet models, we observed an accuracy improvement of 0.35, 1.05, and 1 percentage points respectively for the Flavia database, while 17.42, 3.33, and 11 percentage points respectively for the MalayaKew database.