Case Study: Developing a High-Precision Fungi Classifier for Mobile Deployment

This is a simulated consulting project where I evaluated the feasibility of a mobile application for classifying local fungi as edible or poisonous. Working from the perspective of an ML consultant, I selected and trained models, assessed their performance, and suggested practical strategies for deployment and integration.

The challenge was not just building a model; it was engineering a solution that could run reliably on modest mobile hardware while maintaining life-safety accuracy.

🚀 The Engineering Challenge: Constraints & Performance

Unlike laboratory AI projects, this required a solution designed for the "real world." I had to adhere to strict computational and operational constraints:

• Limited Hardware: Optimized for 4.0 CPU-only environments with a 6.0GB RAM limit.
• Operational Latency:
  • Startup: Initialization and model setup in under 1 minute.
  • Inference: Prediction return in under 200 milliseconds per image for a smooth UX.
• Data Integrity: Developed a custom training and validation pipeline from a dataset of 1,001 labeled images.

🛠️ My Technical Approach

I explored three distinct architectures to find the optimal balance between speed and reliability:

1. ResNet18 (Standard CNN): Achieved 74% accuracy.
2. ResNet18 (Frozen Backbone): Faster training but lower performance at 66% accuracy.
3. DINOv2 (Foundation Model): My recommended algorithm, delivering a significant performance jump to 87.5% accuracy.

Why DINOv2?

By utilizing a Vision Transformer-based foundation model, I was able to leverage self-supervised pre-training to achieve high precision without requiring a massive labeled dataset.

📈 Results & Key Achievements

My final model (DINOv2) delivered metrics that balanced the high stakes of fungi identification:

Metric	Result
Accuracy	87.5%
Precision	92.7%
Recall	89.0%

Handling Data Anomalies

The raw dataset included challenging "unusual data" like drawings, macro-crops, or distant photos. I implemented Data Augmentation (horizontal/vertical flips, rotation, and resizing) to improve robustness. Additionally, I addressed a Class Imbalance where certain species were overrepresented, ensuring the model remained unbiased.

🛡️ Ensuring Reliability & Safety

In software engineering, a model is only as good as its failure handling. I performed a Confidence Calibration analysis which revealed the model is "underconfident"—meaning it often performs better than its internal probability score suggests.

My Recommendations for Integration:

• Dynamic UI Guides: Recommended a mushroom-shaped camera overlay to help users frame shots correctly.
• Environmental Alerts: Logic to trigger "Too Dark" warnings or "One Fungi Per Photo" messages to reduce false negatives.
• Risk Transparency: App displays accuracy rates and clear warnings to ensure users treat the tool as a reference, not a safety guarantee.

💡 Reflection

This project demonstrated my ability to bridge the gap between complex ML research and practical software constraints. By prioritizing latency and resource management, I delivered a high-accuracy solution ready for mobile integration.

Key Skills Utilized:

• Python (PyTorch / PIL)
• Model Optimization for CPU-only environments
• Data Augmentation & Pipeline Engineering
• Reliability & Ethics in AI