基于YOLOv8-MI软枣猕猴桃小目标果实识别和定位方法OA

Soft jujube kiwifruit had rich nutritional value,but due to its small size,dense distribution,and susceptibility to backlighting,the accuracy of fruit identification and positioning during automated harvesting was low,seriously affecting harvesting efficiency.Therefore,a YOLOv8-MI kiwifruit target detection method based on YOLOv8 network structure was proposed.YOLOv8 was optimized,and CBIM enhanced spatial pyramid pooling module was introduced into the backbone network to improve the extraction ability of key features of soft jujube kiwi fruit.The Bi-FPN module was used in the neck network and the small target detection layer was added to enhance the multi-scale feature fusion effect and the small target detection accuracy.The MPDIoU-I loss function was introduced into the head network to dynamically adjust the learning rate to capture the characteristics of small targets and improve the recognition accuracy of fruits under dense occlusion and backlighting conditions.The optimization results showed that the accuracy,recall,and average accuracy of YOLOv8-MI had increased by 8.60,7.50,6.86,and 8.10 percentage point,respectively,while the model weight had only increased by 1.65 MB.Under dense occlusion and backlighting conditions,the accuracy,recall,and average accuracy of the model had increased by 10.28,8.70,and 7.72 percentage point,respectively.Based on the recognition results of YOLOv8-MI,the picking point coordinates were obtained using the SGBM-CL positioning algorithm.Compared with manually cali-brated data,the positioning errors in the X,Y,and Z directions were 9.09 mm,5.98 mm,and 6.10 mm,respectively,which could meet the requirements of picking accuracy.Further identification and positioning verification of the fruit.,the results showed that the overall recognition success rate reached 88%,and the accurate positioning rate reached 82%,which verified the practicality and reliability of the YOLOv8-MI model.