Plant Disease Classification

01 /Overview

Compare building an image classifier from scratch with transfer learning, and measure the difference properly rather than from a single run.

02 /What I built

• Custom CNN: three convolution blocks with batch normalisation and a dense head
• VGG16 pre-trained on ImageNet, frozen, with a new classification head
• Five separate training phases per model for reliable averages
• Training and validation accuracy and loss recorded for every phase

03 /Outcome

VGG16 transfer learning reached 68.11% average validation accuracy against 48.07% for a custom CNN, across five training runs each.

04 /What I learned

Both models overfit (about 99% training accuracy), which made the case for regularisation and more varied data as clearly as the headline comparison.