<p>Addressing the challenges of orange maturity detection in complex natural environments characterized by back-lighting, occlusion, multiple and diminutive targets, this paper proposed a lightweight orange maturity detection model based on an improved YOLOv8.By incorporating an intra-scale feature interaction module based on attention mechanisms (AIFI) into the YOLOv8 network, enhanced the model’s multi-scale feature fusion capabilities while reducing computational complexity and improving detection efficiency and accuracy. Additionally, the VoVGSCSP and GSConv modules replaced the C2f and Conv modules in the feature fusion network, further enhancing the model’s multi-scale feature fusion abilities. Furthermore, a method was proposed for processing and reconstructing orange datasets under natural conditions, which mitigated the interference caused by color variations of orange under different lighting conditions, thereby improving the accuracy of the trained model. Experimental results demonstrated that compared to the original YOLOv3, YOLOv5, and YOLOv8 models based on the raw dataset, the improved algorithm model proposed in this paper achieved improvements in mAP50 of 7.1%, 1.3%, and 2.2% respectively, and average detection time reductions of 19.5%, 38.3%, and 2% respectively. For the improved model based on the reconstructed dataset, mAP50 was improved by 10.8%, average detection time was reduced by 35%, and the number of parameters remained unchanged compared to the improved model based on the raw dataset. The improved model based on the reconstructed dataset achieved a good balance between detection speed, accuracy, and computational complexity. This paper provides guidance for fruit maturity detection in complex environments and serves as a reference for subsequent precision harvesting and intelligent management.</p>