SeedFlow: Sistema de Visão Computacional para classificação de grãos de aveia

Abstract

Oat is a cereal of great importance for human and animal food because of the nutritional benefits it offers from the structures that make up the grain. In practically every step of the production process, the correct identification of the species and the cultivar being used is essential information. The present work establishes a methodology for the acquisition, processing, and classification of digital images of oat grains using computer vision and artificial intelligence techniques. The techniques of deep learning, applied to digital images, are characterized by the use of convolutional neural networks capable of recognizing complex structures present in the acquired images. These techniques were used for two purposes: first, to identify species of oat grains, such as Avena sativa and Avena strigosa, and the second to classify grains for the cultivars of the species Avena sativa. Among the cultivars selected are UPFA Ouro, UPFA Fuerza, and UPFA Gaudéria. Different convolutional network architectures are available in the literature. Thus, six different architectures were compared to identify which would produce the best performance considering the context of this work. This approach provided the 99.7% accuracy result for species identification and the 89.7% accuracy result for the classification of oat cultivars. A computational solution called “SeedFlow” was developed for the use of the proposed methodology. This solution consists of three modules: a software library; an application for training and manipulation of image banks; and an application of grain analysis and classification using pre-trained neural networks. Our approach aims to provide better efficiency compared to the manual methods currently employed. According to the experiment carried out, its use proved to be viable, and it can be a useful tool in pre-selection tests, in laboratory analysis or in making-decision support for plant breeding programs and intellectual property evaluation.