Análise de desempenho da YOLOv5 para detecção em tempo real de plantas daninhas na soja

Abstract

In this work we evaluate a family of You Only Look Once (YOLOv5) object detection models for real-time weed detection in soybean. The images of the plantation were collected in different days, places and climatic conditions, between the phenological stages V3 and V5, considered of high occurrence of weeds in soybean. For training and testing the neural network, we considered four metrics to evaluate classification results: precision, recall, F-measure and accuracy; and three metrics to evaluate the detection results: mean absolute error (MAE), root mean square error (RMSE) and coefficient of determination (R²). The experimental results showed that there is no evidence of a significant difference between the YOLOv5s6/m6/l6/x6 models, but there is evidence of significant difference between the detection scores: 0.25 and 0.35. In our experiments, the detection score of 0.25 performed better for accuracy, recall, and F-measure than 0.35. On the other hand, the detection score of 0.35 was more precision for the detection of weeds in soybean. We also demonstrate how the results generated by detection can be implemented in an end-to-end system in the form of application maps. The results indicate that YOLOv5 can support experts and farmers in weed monitoring to operate in a real field environment.