Logo Lanfrica

samuncleML/SamuncleML

Domaine:

agriculture

Type de record:

model
Créateur:
sam
Hôte:
Multi-Head Plant Disease Classification for African Food Security Using MobileNetV3-Small # Multi-Head Plant Disease Classification for African Food Security ## Overview This project presents a specialized deep learning solution designed to address the **diagnostic deficit** in African agriculture. By leveraging a custom **Multi-Head MobileNetV3-Small** architecture, the system enables **offline, high-precision identification** of crop species and their associated diseases. The goal is to strengthen food security for over **700 million Africans** who depend on staple crops such as **cassava** and **corn**. --- ## Key Features - **Dual-Head Architecture** A shared backbone performs simultaneous: - Plant Identification (6 classes) - Disease Classification (27 classes) - **Mobile-First Design** Optimized for low-resource devices using MobileNetV3-Small with: - Hardware-Aware NAS - Hard-Swish activation - **Out-of-Distribution (OOD) Rejection** Dedicated class to reject non-plant inputs (e.g., soil, tools, hands), preventing invalid predictions. - **Offline Accessibility** Deployed via an offline API to ensure usability in rural regions with limited or no internet connectivity. - **Interpretability** Grad-CAM visualizations highlight disease-relevant regions (lesions, chlorosis), ensuring transparency and trust. --- ## Technical Specifications - **Backbone:** MobileNetV3-Small - **Parameters:** ~2.5 million (≈15× fewer than ResNet-101) - **Input Size:** `224 × 224 × 3` - **Frameworks:** - PyTorch (Training) - TensorFlow Lite (Deployment) ### Accuracy Results - **Plant Identification:** 99.97% Macro F1-score - **Disease Classification:** 98.49% Macro F1-score --- ## Dataset The model was trained on approximately **43,000 images**. To bridge the **lab-to-field gap**, the dataset prioritizes real-world farm images and applies extensive data augmentation, including: - Rotation - Color jittering - Horizontal and vertical flips This improves robustness against inconsistent lighting and background noise. ### Target Crops - Corn (Maize) - Cassava …