MediNetLite: A Resource Efficient Ultra Lightweight CNN for Medicinal Leaf Recognition in Bangladeshi Weed-Infested Area

The main contributions of this work are:

Field-Oriented 34-Class Dataset: To address the key challenge of real-world, cluttered environment medicinal plant recognition, we have introduced a newly compiled dataset of 17050 images captured in the field, comprising 34 medicinal plant species commonly found in weed-infested agricultural areas of Bangladesh.

Custom CNN- MediNet-XG is proposed, which is with 342 KB model size, 98.82% accuracy and 98.81% F1-score which is significantly better than VGG16 (56.20 MB), DenseNet121 (26.98 MB), MobileNetV2 (8.64 MB), and SqueezeNet (889 KB) and is 79×–164× smaller.

Task-Specific Design for Edge Deployment: The architecture utilizes inverted residual blocks with depthwise separable convolutions and lightweight channel attention, achieving both high accuracy and practical deployment on resource-limited devices such as smartphones and edge devices.

Interpretable Multimodal Framework: Integrate Grad-CAM and t-SNE for spatial and feature-space explainability and retrieval-based query module for providing grounded botanical knowledge (uses, safety, dosage) without hallucination for all queries.

Proof-of-concept of Lightweight Superiority: This work demonstrates that in low resource settings, custom ultra-compact, task-specific architectures can outperform generic pretrained models, both in terms of accuracy and efficiency, for fine-grained recognition in agriculture.