Transfer Learning Project: Guava Disease Classification

This project is part of a series exploring the effective use of transfer learning on small datasets. The focus of this project is to demonstrate the importance of using the proper pretrained model weights before fine-tuning for specific tasks. This project focuses on the classification of guava disease types using transfer learning with the VGG19 architecture pretrained on different weights.

Project Workflow

1. Dataset Preparation

The dataset, GuavaDiseaseDataset, is organized into three directories: train, val, and test. Images are resized to 224x224 and augmented using ImageDataGenerator.

Key Steps:

• Dataset Shrinking: To limit dataset size for simulation of limited data scenarios.
• Data Augmentation:
  • Rescaling pixel values.
  • Applying rotations, zooms, flips, and shearing.

2. Model Architectures

Three different pretrained VGG19 models were tested, each fine-tuned for the guava disease classification task.

Model Variants:

1. Imagenet Pretrained VGG19
   Used pretrained weights from ImageNet and added custom dense layers for classification.

2. Fruits and Vegetables VGG19
   A model pretrained on classifying fruits and vegetables , fine-tuned for guava disease classification.

3. Brain Tumor VGG19
   A model pretrained on detecting and classifying brain tumors, fine-tuned for guava disease classification.

3. Training and Fine-Tuning

• Initially, only the added dense layers were trainable, with the pretrained layers frozen.
• Fine-tuning involved unfreezing the pretrained layers and training with a lower learning rate (1e-5).
• Early stopping was employed to prevent overfitting.

4. Evaluation

• Evaluated each model on the test dataset.
• Metrics:
  • Accuracy: Model classification accuracy.
  • Loss: Cross-entropy loss.

5. Visualization

• Plotted accuracy and loss trends using plot_accuracy() and plot_loss().
• Final metrics were summarized with summarize_final_metrics().

Key Python Files

• utils.py: Contains utility functions for dataset shrinking, visualization, and performance summaries.

Results

Each model's performance was compared to determine how well the pretrained weights contributed to guava disease classification.

Metrics:

Model	Accuracy	Loss
Imagenet Pretrained VGG19	95.8%	0.73
Fruits & Vegetables VGG19	67.1%	0.98
Brain Tumor VGG19	91.8%	0.91

Findings:

The architecture for all those models was the exact same, taking VGG19 as the base and simply changing the starting weights.

• The Imagenet VGG19 model provides the best performance, being a great multipurpose pretrained

• model that is easy to fine tune on different datasets.

• The Fruits and Vegetables VGG19 model, although pretrained on classifying fruits and vegetables, did not perform as well as the other models. This might be due to how much it relied on higher level features to classify fruits, focusing more on less detailed information.

• The Brain Tumor VGG19 model, although pretrained on detecting and classifying brain tumors, performed quite well. Although this is an odd choice of pretrained weights for guava disease classification, but it focuses more on finding anomolies in more detailed information, taking lower layers into stronger account. This explains the surprisingly good performance on our mode.