Abstract: In order to achieve precise and fast identification of citrus fruits, an improved YOLOv8 was proposed. Firstly, certain traditional convolutions in the YOLOv8 were replaced with ODConv full-dimensional dynamic convolutions to enhance the model’s adaptability in complex natural environments. Subsequently, the CIoU loss function of YOLOv8 was substituted with the MPDIoU loss function to address the degradation issue of the CIoU loss function in specific scenarios. Furthermore, the effectiveness of ODConv full-dimensional dynamic convolutions and MPDIoU loss function was verified through a series of ablation experiments. The average recognition accuracy （mAP） of the improved models, YOLOv8n, YOLOv8s, YOLOv8m, YOLOv8l, and YOLOv8x, was increased from 86.40%, 88.92%, 88.97%, 88.99%, 89.11% to 88.25%, 89.32%, 89.57%, 89.90%, 90.12%, respectively. Experimental results demonstrated that ODConv full-dimensional dynamic convolutions and MPDIoU loss function significantly enhanced the citrus fruit identification capability of the YOLOv8 in natural environments.

Key words: citrus fruit recognition, convolutional neural network, YOLOv8, ODConv full-dimensional dynamic convolution, MPDIoU loss function