Deep Sporenet: A Lightweight Few-Shot CNN For Illumination-Robust Fungal Species Identification on Mobile Devices

Abstract

Fungal infections in plants and humans pose major challenges to food security and clinical diagnostics, yet their identification still depends on expert microscopy and culture-based methods that are slow and equipment-intensive. Existing deep learning models for fungal classification achieve high accuracy under controlled imaging conditions but fail under illumination shifts, class imbalance, and low-data regimes typical of field or point-of-care scenarios. This study introduces Deep SporeNet, a compact convolutional neural network (CNN) with a few-shot episodic learning head and illumination-robust preprocessing, designed for mobile deployment in agricultural and clinical environments.
The proposed framework integrates (i) color constancy and stain normalization to counter variable lighting, (ii) a MobileNetV2/EfficientNet-Lite backbone for efficient feature extraction, (iii) a Prototypical Network head for low-sample fungal taxa recognition, (iv) entropy-based test-time adaptation (TENT) for on-device robustness, and (v) temperature scaling for confidence calibration. Evaluations on the MycoAI-Lab and FieldMyco-Real datasets demonstrate that Deep SporeNet achieves 94.2% accuracy, 92.8% macro-F1, and a tail-class F1 of 82.7%, outperforming state-of-the-art mobile CNNs while running in < 50 ms on a standard smartphone processor. Its well-calibrated predictions (ECE = 0.039) and interpretable Grad-CAM visualizations confirm suitability for real-time fungal diagnostics, crop protection, and biodiversity monitoring. The model thus represents a scalable step toward AI-assisted mycology that is both data-efficient and deployable in resource-limited settings.