Target Detection of Green-Skinned Walnut from the Perspective of UAV Based on Improved YOLO v8

At present, there is relatively little research on walnut data as the target detection object. Secondly, the existing walnut data sources are mostly handheld devices or ground collection devices. Moreover, most of the methods for walnut production measurement still rely on traditional statistical models, and their accuracy cannot be guaranteed. Finally, although some other papers have used deep learning methods, their real-time detection performance is extremely low. In order to improve the accuracy and real-time detection of green skinned walnuts from the perspective of drone aerial photography, we have established a walnut image dataset. For the first time, an efficient medical image segmentation network based on large kernel and skip fusion was integrated into the YOLO v8 model, and a multi-scale feature dynamic convolution kernel was designed. Finally, the improved YOLO v8 CM model was used for object detection of green skinned walnuts. The experimental results show that the improved YOLO v8-CM model has improved precision, recall, and mAP50 values by 2.0%, 1.1%, and 1.0%, respectively, compared to the original YOLO v8 model. Moreover, the above indicators have improved by 47.9%, 13.3%, and 34.8% respectively compared to Faster R-CNN, and its detection speed is 7.3 times faster than Faster R-CNN.