Scholarlink

<p>Identification and classification of Ayurvedic plants are crucial for pharmaceutical research, traditional medicine, and biodiversity conservation. Manual classification is characterised by its labour intensity, susceptibility to errors, and the necessity for specialised expertise. Existing deep learning models often struggle with high computational costs and the challenge of distinguishing between species with similar appearances (inter-class similarity). In this study, we propose MediPlantNet, a novel dual-backbone, feature-level fusion ensemble method. MediPlantNet integrates modified MobileNetV4-Small and EfficientNetV2-B0 architectures to enhance classification accuracy while significantly reducing trainable parameters and computational complexity, measured in Giga Floating-Point Operations per Second (GFLOP). By concatenating feature vectors from both models and utilizing advanced data augmentation, our model robustly learns distinctive plant characteristics. On a comprehensive dataset of 80 medicinal plant species, MediPlantNet achieves state-of-the-art performance with 99.32% accuracy, 99.36% precision, 99.34% recall, and a 99.33% F1-score, with a perfect Area Under the Curve (AUC) of 1.00. A comparative analysis confirms that MediPlantNet outperforms models such as ResNet101, InceptionV3, and VGG19, offering superior performance with fewer parameters (10.1M) and lower GFLOP (0.841). This work demonstrates the efficacy of our feature-fusion approach for accurate, efficient, and scalable automated botanical classification.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Mediplantnet: a dual-backbone feature fusion method for high-precision classification of medicinal plant species

  • 期刊论文

  • 发表日期：

    2026年1月7日

查看全文
  • Dheeraj Kumar,
  • Manish Kumar Sharma,
  • Piyush Kumar Singh,
  • Prabhat Ranjan

摘要

Identification and classification of Ayurvedic plants are crucial for pharmaceutical research, traditional medicine, and biodiversity conservation. Manual classification is characterised by its labour intensity, susceptibility to errors, and the necessity for specialised expertise. Existing deep learning models often struggle with high computational costs and the challenge of distinguishing between species with similar appearances (inter-class similarity). In this study, we propose MediPlantNet, a novel dual-backbone, feature-level fusion ensemble method. MediPlantNet integrates modified MobileNetV4-Small and EfficientNetV2-B0 architectures to enhance classification accuracy while significantly reducing trainable parameters and computational complexity, measured in Giga Floating-Point Operations per Second (GFLOP). By concatenating feature vectors from both models and utilizing advanced data augmentation, our model robustly learns distinctive plant characteristics. On a comprehensive dataset of 80 medicinal plant species, MediPlantNet achieves state-of-the-art performance with 99.32% accuracy, 99.36% precision, 99.34% recall, and a 99.33% F1-score, with a perfect Area Under the Curve (AUC) of 1.00. A comparative analysis confirms that MediPlantNet outperforms models such as ResNet101, InceptionV3, and VGG19, offering superior performance with fewer parameters (10.1M) and lower GFLOP (0.841). This work demonstrates the efficacy of our feature-fusion approach for accurate, efficient, and scalable automated botanical classification.