AM-YOLO: An Efficient Object Detection Model for <i>Astragalus mongholicus</i> Authentication With Dynamic Gated Feature Fusion

Astragalus membranaceus (AM) holds significant medicinal value in clinical applications. However, the market frequently encounters phenomena where morphologically similar counterfeits are used to impersonate AM, severely affecting the efficacy of medications and drug safety. To achieve rapid and accurate identification of genuine and counterfeit AM, this paper proposes an improved object detection model based on YOLOv8—AM-YOLO. The model achieves adaptive weighted integration of high- and low-level features by designing a gating fusion module L2Hblock. It proposes an improved spatial pyramid pooling structure ESPPF to enhance local saliency and global context awareness. Additionally, the lightweight Faster-Block is adopted to construct the Fc2f module, reducing the model’s computational complexity. Additionally, the introduction of a dynamic upsampling module, DySample, further improves the ability to restore edge details. To validate the effectiveness of AM-YOLO, we constructed the first AM detection dataset, covering 11 common fakes. The experimental results show that AM-YOLO achieves 95.7% in the mAP50-95 metric, an increase of 2.8% compared to YOLOv8, with an inference speed of 176 Frames Per Second (FPS). Compared with the state-of-the-art object detection algorithms, AM-YOLO reduces the average error rate by 30%. Moreover, the model can maintain leading inference speed when deployed on devices with different computing power. The datasets, code and test results used in this study are available https://github.com/feifeimao888/AM-YOLO.