Intelligent Detection of Northern Shaanxi Red Fuji Apples based on Improved YOLOv8 Model

Abstract

Detection of Northern Shaanxi Red Fuji apples in natural orchards faces challenges such as branch-leaf occlusion, fruit overlapping, illumination variation, and scale differences, leading to decreased detection accuracy and localization stability of the model. To address these issues, YOLOv8s was employed as the baseline model, with improvements made from three aspects: small object detection layer, attention mechanism, and bounding box regression loss. Specifically, a P2 detection layer, Coordinate Attention, EIoU, and WIoU were introduced. Multiple ablation experiments were conducted under unified training conditions, and best.pt was adopted for re-validation to avoid deviations caused by early stopping. Experimental results demonstrate that among the improved schemes, YOLOv8s+EIoU achieved the optimal performance, with mAP50-95 increased to 69.73%, representing a 0.21 percentage point improvement over the baseline, which verifies the optimization effect of EIoU on bounding box regression. This study provides technical references for intelligent apple recognition, automated harvesting, and orchard management.