GCD-YOLO: A deep learning network for accurate tomato fruit stalks identification in unstructured environments

• Simultaneous detection of fruit stalks and determination of fruit ripeness. • A tomato fruit stalks detection method based on GCD-YOLO was proposed. • Efficient deployment of the GCD-YOLO model on NVIDIA Jetson Orin Nano was achieved. Accurate identification of tomato fruit stalks is critical for end-effectors to perform precise cutting operations, which contributes to enhanced harvesting efficiency and reduced reliance on manual labor. However, actual field environments may compromise the accuracy of tomato fruit stalks identification. To address this issue, this study proposes an improved GCD-YOLO model based on YOLOv8n. Specifically, GAM is integrated into the YOLOv8 backbone network, the original Neck is replaced with CCFM, and Dyhead is adopted as the detection head for tomato fruit stalks identification. The results show that the different improvement modules can effectively enhance the feature extraction ability of the model, thus improving the detection accuracy of the GCD-YOLO model for tomato fruit stalks. The GCD-YOLO model achieves a detection precision of 94.4% and an mAP@50 of 91.7% for tomato fruit stalks, significantly outperforming YOLOv8n, SSD, Faster R-CNN, YOLOv5n, RT-DETR, YOLOv9t, YOLOv11n and YOLOv12n. Furthermore, the GCD-YOLO model was deployed on an NVIDIA Jetson Orin Nano for field experiments. Experimental results demonstrate that the GCD-YOLO model delivers outstanding detection performance on edge devices, achieving an inference speed of 26.24FPS in actual field environments. The research findings contribute to the development of smart agriculture and facilitate the implementation of machine vision in tomato harvesting applications.