Optimizing edge-enabled system for detecting green passion fruits in complex natural orchards using lightweight deep learning model

To address labor shortages and rising costs, developing cost-effective fruit detection technology capable of functioning effectively in complex orchard environments is especially crucial for the advancement of robotic passion fruit harvesting systems. Moreover, achieving edge device-based efficient detection is highly expected under field conditions given its operating portability and cost-effective effects. This study proposed an improved YOLOv8n model for automatic passion fruits detection. First of all, a ParNet attention mechanism was added to the C2f module of YOLOv8n to improve feature extraction. To extract more information about small targets in the images, an additional detection layer was added for small targets in the Neck network. Furthermore, a SlimNeck architecture was employed to optimize the original neck part, reducing the model parameters while maintaining detection performance. The proposed model was trained and tested using a dataset divided by Hold-out, achieving an accuracy of 96.0 %, a recall rate of 83.7 %, and a [email protected] of 91.9 %. The model size was optimal with 2,650,300 parameters, 10.4G FLOPs, and an inference speed of 115fps in Windows-based platform. Compared to the other state-of-the-art deep learning models such as YOLOv4-Tiny, YOLOv5n, YOLOv6n, YOLOv7-Tiny, YOLOv8n, YOLOv9, YOLOv10n, YOLOv11n, Faster R-CNN and SSD, the improved YOLOv8n model showcased overall superior detection performance. When deploying this proposed model on Nvidia Jetson Orin Nano, the inference speed of the improved model was 28.15fps in the C++ environment using the TensorRT API, showing real-time detection performance. This study can provide basic technology for passion fruit robotic harvesting on the basis of the potable edge devices.