ScaleEdgeFusion-Network: a lightweight and efficient model for citrus maturity detection in complex orchards

Abstract

Accurate detection of citrus fruit maturity is essential for robotic selective harvesting in orchards yet remains challenging due to complex environmental conditions. This paper presents ScaleEdgeFusion-Net (SEF-Net), a lightweight object detection framework built on YOLO11n for efficient citrus maturity recognition in real-world orchard environments, achieved through three key innovations: an Adaptive Multi-scale Edge Enhancement module integrated with the backbone to improve fruit discriminability; an Enhanced Multi-scale Feature Extraction module replacing standard spatial pyramid pooling to strengthen robustness against complex environmental conditions; and an Efficient Feature Fusion and Dynamic Sampling Neck redesigned to leverage dynamic upsampling and channel attention for high detection accuracy with minimal computational overhead. Experimental results demonstrate that SEF-Net achieves superior performance with 93.0% mAP@0.5 while maintaining only 2.0 million parameters and 5.2 G, resulting in a compact model size of 4.3 MB, and delivers the highest inference speed (132 FPS) among all compared models. Compared to state-of-the-art detectors—including general-purpose models (YOLOv5n, YOLOv8n, YOLO11n, etc) and citrus-specific models such as ORD-YOLO and LightSal-DETR—the proposed method achieves higher detection accuracy with significantly lower computational requirements. These results indicate that SEF-Net provides an effective balance between accuracy and efficiency, making it suitable for deployment on resource-constrained harvesting robots in precision agriculture applications.