Hi Everyone 
I wanted to share a computer vision app I recently built using Streamlit + PyTorch that detects common apple leaf diseases from a single photo and provides tailored care recommendations for each predicted condition.
Try the app here
Github repository
What’s under the hood?
-
Fine-tuned ResNet-18
-
Calibrated probability outputs (temperature scaling)
-
TorchScript export for fast, lightweight CPU inference
-
Deterministic preprocessing pipeline (resize → center crop → normalize)
-
Confidence threshold routing to an
"unknown"label when the model isn’t confident
The model classifies leaf images into:
-
Healthy
-
Scab
-
Rust
-
Black Rot
If the prediction confidence is below a defined threshold, it safely returns “unknown” instead of over-committing. I wanted to keep it practical and conservative for real-world usage.
Smart input handling:
The app supports:
-
Image upload
-
Live camera capture
For camera images, I added:
-
Brightness validation
-
Leaf-likeness gating (green coverage + texture checks)
-
Optional bypass for edge cases
This helps reduce false positives and noisy captures before inference even runs.
Tailored Recommendations:
After prediction, the app doesn’t just stop at a label.
Each class triggers specific, actionable care guidance:
-
Airflow improvement strategies
-
Fungicide timing suggestions
-
Irrigation best practices
-
Host management considerations
So the output becomes a decision-support tool, not just a classifier.
Built With:
-
Streamlit (UI + interaction layer)
-
PyTorch
-
Torchvision
-
TorchScript (for production-ready model deployment)
-
Custom probability rendering and structured class mapping
Everything runs locally on CPU, so no cloud dependency.
Would love feedback from the community, especially around:
-
UI/UX improvements
-
Model calibration strategies
-
Additional validation layers
-
Ideas for expanding to other crops
Appreciate any thoughts!