YOLO-TMHC: An Optimized Model Balancing Precision and Efficiency for Tibetan Medicinal Herb Recognition
Year: 2025
Type: conference-paper
Abstract: Accurate recognition of Tibetan medicinal herbs is crucial for the research and application of modern traditional Chinese medicine. However, existing manual recognition methods face challenges such as high labor intensity and high misidentification rates. Therefore, this study proposes a deep learning-based classification model for Tibetan medicinal herbs, named YOLO-TMHC. YOLO-TMHC is built upon YOLOv11n and incorporates various improvement strategies to meet the high accuracy requirements of practical applications. Specifically, the C3k2 module of the backbone network is improved by adding a concatenated diverse branch block (DBB) module. By combining multi-scale convolution, sequential convolution, and average pooling operations, the model is able to learn more comprehensive and specific target features. After the connection operation in the neck structure, an efficient bidirectional feature pyramid network (BiFPN) is introduced to achieve bidirectional fusion and weighted optimization of multi-scale features, further enhancing the model's ability to extract key features in object detection. Additionally, a CAFM attention mechanism is added after each C3k2 module in the neck to further improve the model's learning of key features. This study uses a dataset based on 12 categories of Tibetan medicinal herbs, consisting of 29,384 images. Experimental results demonstrate that the proposed YOLO-TMHC achieves the best recognition performance on this herbal dataset. The [email protected] of YOLO-TMHC is 98.14%, which represents a 1.32% improvement over YOLOv11n in [email protected]. Moreover, YOLO-TMHC recognizes a single image in just 28.6 ms. The experimental results show that YOLO-TMHC outperforms currently mainstream recognition models and achieves high-precision recognition of Tibetan medicinal herbs. This study provides technical support for the development of modern herbal recognition and edge recognition devices.
Authors: Xinlong Hu, Yue Shen, Nuo Qun, Ji De
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Awards: G2275401422, G8622653477
OpenAlex ID: W7127046222