Bean Disease Classification

Stepwise CNN fine-tuning achieving 98.86% test accuracy — MSDS coursework research

Research Paper MSDS Coursework 2025

“Improving Bean Disease Classification with Stepwise Fine-tuning of CNN-based Transfer Learning Model”

Overview

Benchmarked 5 pretrained CNN architectures on a combined bean leaf disease dataset and identified ResNet50 as the optimal backbone through systematic comparative evaluation.

Architectures Benchmarked

  • VGG16
  • ResNet50 (selected as optimal backbone)
  • DenseNet121
  • EfficientNet-B0
  • MobileNetV2

Stepwise Fine-Tuning Strategy

Designed and implemented a stepwise fine-tuning strategy — progressively unfreezing layers to adapt pretrained ImageNet representations to the plant disease domain — achieving 98.86% test accuracy on the final model.

Pipeline

Conducted an end-to-end pipeline: data loading, preprocessing, multi-model benchmarking, fine-tuning, evaluation, and visualization in a single reproducible Kaggle notebook.

Tech Stack: Python · TensorFlow/Keras · Kaggle GPU · Jupyter