Resumen
Due to differences in planting environment, color, shape, size, and compactness, accurate detection of grape clusters is very difficult. Herein, a real-time detection method for grape clusters based on the YOLOv5s deep learning algorithm was proposed. More specifically, a novel dataset called Grape-internet was constructed, which consisted of 8657 grape images and corresponding annotation files in complex scenes. By training and adjusting the parameters of the YOLOv5s model on the data set, and by reducing the depth and width of the network, the lightweight processing of the network was completed, losing only a small amount of accuracy. As a result, the fast and accurate detection of grape clusters was finally realized. The test results showed that the precision, recall, mAP and F1 of the grape cluster detection network were 99.40%, 99.40%, 99.40% and 99.40%, respectively, and the average detection speed per image was 344.83 fps, with a model size of 13.67 MB. Compared with the YOLOv5x, ScaledYOLOv4-CSP and YOLOv3 models, the precision of YOLOv5s was 1.84% higher than that of ScaledYOLOv4-CSP, and the recall rate and mAP were slightly lower than three networks by 0.1?0.3%. The speed was the fastest (4.6 times, 2.83 times and 6.7 times of YOLOv3, ScaledYOLOv4-CSP and YOLOv5x network, respectively) and the network scale was the smallest (1.61%, 6.81% and 8.28% of YOLOv3, ScaledYOLOv4-CSP YOLOv5x, respectively) for YOLOv5s. Moreover, the detection precision and recall rate of YOLOv5s was 26.14% and 30.96% higher, respectively, than those of Mask R-CNN. Further, it exhibited more lightweight and better real-time performance. In short, the detection network can not only meet the requirements of being a high precision, high speed and lightweight solution for grape cluster detection, but also it can adapt to differences between products and complex environmental interference, possessing strong robustness, generalization, and real-time adaptability.