PlantScan AI
Deep-learning powered plant disease diagnosis from a leaf image, with model comparison and a camera-ready Flask application.
Three CNN backbones were trained and evaluated. EfficientNetB0 and MobileNetV2 share the same 38-class dataset and are the interchangeable models exposed by the application.
A leaf image should be enough to start a diagnosis.
Plant disease identification can be slow when a grower has to rely on manual inspection or specialist access. PlantScan AI turns a leaf photo into a model-backed diagnosis report with confidence, symptoms, treatment and prevention information.
Transfer learning across three CNN backbones.
The project uses ImageNet-pretrained CNN backbones and a two-phase training recipe: first train a new classification head with the backbone frozen, then unfreeze the last 30 layers and fine-tune at a low learning rate. Images are resized to 224×224 with augmentation and an 80/20 train-validation split.
38-class PlantVillage-color model; 98% final accuracy and the app's default model.
38-class PlantVillage-color model; 95% final accuracy and a smaller/faster alternative.
96% on a separate 15-class PlantVillage subset; not directly comparable to the other two.
The numbers are only useful when the datasets match.
| Model | Dataset | Classes | Params | Accuracy |
|---|---|---|---|---|
| EfficientNetB0 | PlantVillage-color | 38 | 4.1M | 98% |
| MobileNetV2 | PlantVillage-color | 38 | 2.3M | 95% |
| ResNet50 | PlantVillage subset | 15 | 23.6M | 96% |
ResNet50 was trained on a different, smaller dataset containing 15 classes from Pepper, Potato and Tomato. Its 96% result should therefore not be presented as a direct head-to-head win or loss against the 38-class EfficientNetB0 and MobileNetV2 models.
From camera frame to diagnosis report.
The live camera path uses the browser camera API and canvas capture, then sends the captured frame through the same inference pipeline as an uploaded image. A lightweight browser-side quality check can auto-capture once the frame is well-lit, sharp and steady.
Why the production dropdown only exposes two models.
EfficientNetB0 and MobileNetV2 were trained on the same 38-class dataset and share the same class mapping, so they are safe interchangeable choices in the application. ResNet50 is intentionally excluded from that selector because its 15-class dataset uses a different label mapping; wiring it into the same selector without separate handling could produce incorrect labels.
EfficientNetB0 because it reached 98% on the shared 38-class evaluation.
MobileNetV2 at 95%, useful when a smaller/faster backbone is preferred.
ResNet50 remains part of the research comparison, not the shared production selector.
A compact computer-vision stack from training to inference.
98% on the shared 38-class evaluation.
The model comparison is only the beginning.
Explore the training notebooks, Flask inference service, camera workflow and full model-comparison report in the repository.