Sistema para classificação automática de grãos de soja baseado em visão computacional

Abstract

The classification of soybean grains for quality control in agribusiness is an essential task, traditionally performed by human classifiers, involving the separation and visual analysis of the grains. However, this process is time-consuming and prone to human errors. In this work, we address the automation of this procedure through the analysis of grain images using computer vision techniques. To build an automated classification system, we tested various approaches and neural network architectures. Firstly, we created a dataset containing images of soybean grains of various classes, particularly focusing on severe damage classes: scorched, moldy, and burnt grains. We used the U-Net architecture for semantic segmentation, followed by digital image processing techniques to isolate the damaged grains. An alternative based on grain detection by a YOLO V4 network was also tested. Additionally, we evaluated convolutional networks such as EfficientNet B1 and MobileNet V1 for damage classification. Our experiments indicate that a cascade approach, with grain detection/segmentation followed by a classification step, yielded superior results to an approach with a single stage of grain detection already with their classes. The approach based on the YOLO V4 network for detection, followed by MobileNet V1 for classification, achieved an IoU value above 0.77 for the separation between healthy and damaged grains, and an accuracy of 0.8926 in damage classification, with a sensitivity of 0.8913 in identifying burnt grains — a metric that should be noted, as these grains can spontaneously combust, causing accidents and silo loss. Balancing between result quality and processing speed, this approach was chosen to be incorporated into a system for soybean grain classification produced by the company Neosilos, which has been marketed to sector cooperatives.