What you build
A convolutional neural network that classifies leaf photos as healthy or as one of several diseases, trained on a public plant-disease dataset, plus a simple interface where a user uploads a photo and gets the result.
Modules
- Dataset preparation and augmentation
- Transfer learning with a pre-trained CNN
- Training, validation and testing
- Upload-a-photo interface with the predicted disease and confidence
What you will understand
- How a CNN learns features from images
- Why transfer learning works with small datasets
- Overfitting and how to detect it
How it is tested
- Accuracy, precision and recall per class
- Confusion matrix
- Testing on photos taken outside the dataset
What goes in the report
- Literature survey
- Dataset description and preprocessing
- Model architecture, results and limitations
Viva questions to prepare
- Why did accuracy drop on your own photos?
- What does data augmentation do?
- Explain one layer of your model.