Apple Recognition Method in Orchards Based on SAM and YOLOv8

The aim of this research is to improve automatic harvesting of orchard apples through an efficient detection method. By applying TensorRT, the YOLOv8 model will run much more efficiently while optimising computational resources. In particular, we believe that the depth and complexity of various YOLOv8 versions of the model will play a key role in improving the detection performance. Therefore, in this study, we tested several versions of YOLOv8 algorithmic recognition models, such as YOLOv8s, YOLOv8n, YOLOv8l, YOLOv8m, etc., and used a variety of model annotation methods including traditional manual annotation, unsupervised annotation, and semiautomatic annotation tools based on the large-scale SAM model SAMsaa. In addition, we tested the effectiveness of automatic apple detection in orchards with and without hardware acceleration. In order to test the above hypotheses, we conducted several experiments and showed that the overall detection performance of the YOLOv8m model was significantly improved in the experimental setting where the dataset was labelled using the SAMsaa tool and optimised using TensorRT. In addition, the overall detection performance of the YOLOv8m model was even more significantly improved in the experiments where the TensorRT-optimised dataset was labelled using the SAMsaa tool on the Jetson Xavier computing platform. The detection mAP50 improved by 33% and 32.7%, respectively, and the average detection accuracy for apple detection reached 90.41%. These results validate the effectiveness and superiority of our method.