Detection of maturity of “Okubo” peach fruits based on inverted residual mobile block and asymptotic feature pyramid network

Abstract

The harvesting of “Okubo” peach fruits is important in food processing and requires intelligent detection. In this study, a lightweight detection model YOLO-IA is proposed based on YOLOv8s combined with an inverted residual mobile block (iRMB) and asymptotic feature pyramid network (AFPN). Firstly, the C2f_iRMB module is designed to replace all C2f modules of YOLOv8s by using the iRMB module, which improves the model’s ability to extract features and detect accuracy. Secondly, the AFPN feature fusion method is adopted for the neck network to enhance the fusion ability of the model to the features of the backbone network, optimize the model parameters, and realize the model’s lightweight. Finally, the “Okubo” peach fruit detection system was developed, which can detect the fruit information in real-time. The results show that the YOLO-IA model has an average precision (AP) of 93.17% and 95.63% for unripe and ripe peaches, respectively, and the mean average precision (mAP) of 94.40%, with a model size of 10.9 MB and an inference time of 6.3 ms, which is capable of real-time detection. Compared with YOLOv8s, YOLO-IA improved the mAP and F1 scores of “Okubo” peach fruits by 1.38% and 1.53%, respectively, and compressed the model size by 49.07%. In summary, YOLO-IA is effective in detecting “Okubo” peach fruits in complex orchard environments and can provide a theoretical basis for the development of subsequent vision systems for picking robots.