Design and implementation of a lightweight real-time detection model (GCA-YOLOv8n) for apple ripeness

Objective  To address issues with traditional apple ripeness detection models, including excessive size, slow inference speed and low detection accuracy.

Method We constructed a lightweight real-time detection model, GCA-YOLOv8n, based on an improved YOLOv8n. First, the C3Ghost module replaced the original model’s C2f module to achieve lightweight design and enhance inference speed. Second, the GhostConv module substituted the original Conv layer to improve information extraction efficiency and reduce redundancy in convolutional layers. Finally, the ACmix attention mechanism was integrated into the original model architecture to boost feature extraction capability and detection accuracy. The GCA-YOLOv8n model was then applied to apple ripeness detection experiments.

Result Experimental results showed that the GCA-YOLOv8n model achieved 2.0×10⁶ parameters, 5.7×10⁹ floating point operations, and a weight file size of 4.4 MB, representing reductions of 33.1%, 29.6%, and 30.2% respectively compared to YOLOv8n. The inference speed reached 130.8 frames per second, a 21.5% improvement over YOLOv8n. The mean average precision and F1 score were 89.2% and 82.5% respectively, demonstrating high detection accuracy and inference speed.

Conclusion The constructed GCA-YOLOv8n model significantly reduces model complexity and computational load while maintaining detection accuracy, achieving lightweight and efficient performance. The model demonstrates high and real-time detection capability and can run stably on mobile edge devices, providing technical support for automated harvesting.