↓Skip to main content

Bean Disease Classification

Table of Contents

Research Paper | MSDS Coursework | Batangas State University | 2025
Improving Bean Disease Classification with Stepwise Fine-Tuning of CNN-based Transfer Learning Models
GitHub Repository →

Overview

Agricultural computer vision frequently encounters high intraclass visual variation and limited domain-specific labeled samples. This project benchmarks 5 state-of-the-art pretrained CNN architectures on a composite bean leaf disease dataset, identifying ResNet50 as the optimal backbone and enhancing its performance through progressive layer unfreezing.

View GitHub Code

Architectures Benchmarked

  • VGG16
  • ResNet50 (Selected as highest performing backbone)
  • DenseNet121
  • EfficientNet-B0
  • MobileNetV2

Stepwise Fine-Tuning Methodology

Rather than applying naive one-shot transfer learning, we designed a stepwise fine-tuning strategy:

  1. Feature Extraction: Frozen base weights with freshly trained classification heads.
  2. Top-Block Unfreezing: Selective fine-tuning of upper residual blocks with reduced learning rates.
  3. Deep Residual Adaptation: Gradual unfreezing of deeper convolutional stages paired with cosine decay scheduling, mitigating catastrophic forgetting.

This systematic adaptation drove the final model to 98.86% test classification accuracy.

Reproducibility

The entire pipeline—data augmentation, class rebalancing, multi-model evaluation, confusion matrix generation, and Grad-CAM visualization—was engineered as a completely self-contained, reproducible notebook on Kaggle GPU.

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