Utilização de comitês de Redes Neurais Artificiais na classificação de danos em sementes de girassol

Abstract

Brazil has the third largest world seed market and invoices billions of reais annually, making seed analysis extremely important. Through analysis techniques, it is possible to determine the germination potential and identify damage to the seeds. X-ray is one of the most desirable techniques, because it provides fast analysis and does not result in the destruction of seeds. However, the images resulting from the X-ray process often require post-processing, seeking visual improvement of the images for analysis by the evaluators. The evaluation process can be carried out by one or more evaluators, but it has a lot of subjectivity, making the automation of this analysis interesting. ANNs (Artificial Neural Networks) are known to be effective for use in pattern recognition and data classification problems, making them good candidates for this automation. In this work, the goal was to perform the classification of radiographic images of sunflower seeds, according to their level of damage, using multiple techniques of image processing and extraction of characteristics to compose different datasets in order to train the ANNs. For this, a dataset consisting of radiographic images of sunflower seeds was used, classified by evaluators into three categories: filled, partially filled or deformed seeds. Using these images, datasets were composed and used to train, validate and test ANNs, which were then used to compose committees. For each case, 10 committees were formed, and obtained averages of the metrics of accuracy, AUC and Kappa index of the committees. The averages of the performance metrics, approximately 90% for accuracy, 0.97 for AUC and 0.84 for Kappa, in the best case, indicate the efficiency of the methodology used in this work and suggest the possibility of using it in composition to usual methodologies for seed classification and evaluation.