GinkgoSense-Net: A method for characterization and real-time detection of ginkgo fruits in complex field environments

Automatic picking of ginkgo fruits is beneficial for prolonging their freshness and ensuring quality during storage. Accurate identification of ginkgo fruits is a critical prerequisite for enabling stable and reliable operation of the end-effector during the harvesting process. However, due to the similarities in color and size between ginkgo fruits and leaves, traditional recognition methods are prone to being affected by variations in lighting, occlusion, and background interference. To solve these challenges, this paper presents an accurate identification method for ginkgo fruits based on GinkgoSense-Net. Specifically, EMA and SAF are integrated into the YOLOv11n framework, and the bounding box regression function is replaced with EIoU. At the same time, methods such as Grad-CAM visualization, ablation experiments, and comparative experiments are used to validate the efficacy of GinkgoSense-Net. The experimental results demonstrate that GinkgoSense-Net has a precision of 95.0%, a recall of 92.1%, and an mAP of 93.6% in the detection of ginkgo fruits, and its real-time performance reaches 74.9. These findings indicate that the proposed model achieves an effective balance among detection accuracy, computational efficiency, and model complexity. The research findings facilitate the advancement of smart agriculture and promote the practical application of vision technology in the field of intelligent harvesting.