Comparative study of citrus fruits (Citrus reticulata Blanco cv. Batu 55) detection and counting with single and double labels based on convolutional neural network using YOLOv7

Abstract

• Performance of different YOLOv7 architecture of deep learning model has been compared. • The model has been trained and tested for study case of citrus fruit detection and counting. • Original YOLOv7 shows the best performance compared to tiny and x version. Fruits detection and counting is an important task for yield prediction that could be achieved by computer vision. The ability to locate and count the fruits could also help the harvesting robot to do a picking task. YOLO is one of the deep learning models which is popular and widely used for object detection and has good performance in detection speed and precision. In the citrus counting task, the label could be set as a single label or multi-label which shows different citrus maturity. The performance of the deep learning model could be different with a different number of labels. Furthermore, there are several types of YOLOv7 models with different sizes and purposes which could also have different performances to do a similar task. This study aims to compare the performance of different kinds of YOLOv7-based deep learning models for citrus fruit detection and counting. Case study on the citrus cv. Batu 55 trees have been carried out. The results show that the original YOLOv7 achieved the best performance both on single and double labels compared to the tiny and X versions of YOLOv7. The YOLOv7 could reach mAP50 of 0.906, a precision of 0.85, a sensitivity of 0.825, and an F1-score of 0.837, while for the counting task, the model has a good performance with R 2 of 0.966.