Optimizing Image Classification Model: Exploring with CNNs, Transfer Learning, Hyperparameter Tuning, and K-Fold Cross-Validation

This project focuses on optimizing an image classification model using Convolutional Neural Networks (CNNs), transfer learning, hyperparameter tuning, and K-Fold Cross-Validation. The goal is to explore various techniques to improve model performance on the CIFAR-10 dataset.

Disclaimer

This project is for exploration and testing purposes only and is not intended for production use.

Introduction

The CIFAR-10 dataset consists of 60,000 32x32 color images in 10 classes, with 6,000 images per class. The project aims to build and optimize a CNN model to classify these images accurately.

Dataset

The CIFAR-10 dataset is loaded and preprocessed using the following steps:

• Normalization of pixel values to be between 0 and 1.
• Conversion of labels to one-hot encoding.

Methodology

Data Augmentation

Enhanced data augmentation techniques are applied to improve model generalization. These include:

• Width and height shifts.
• Horizontal flips.
• Rotations.
• Zooming.
• Shearing.

Model Architecture

A slightly more complex CNN model is built with the following layers:

• Convolutional layers with batch normalization.
• MaxPooling layers.
• Dense layers with dropout for regularization.

Hyperparameter Tuning

Hyperparameters are tuned using Optuna, an automatic hyperparameter optimization framework. The following hyperparameters are optimized:

• Learning rate.
• Batch size.

Transfer Learning

Transfer learning is applied using pre-trained models such as MobileNetV2 and VGG16. The base models are frozen, and custom top layers are added for classification.

K-Fold Cross-Validation

K-Fold Cross-Validation is used to evaluate the model's performance more robustly. The dataset is split into 2 folds, and the model is trained and validated on each fold.

Results

Training and Validation Curves

The training and validation curves show the model's performance over epochs. The plots indicate the accuracy and loss trends.

Hyperparameter Optimization

The best trial from Optuna optimization yielded the following parameters:

• Learning rate: 0.0008750803832793538
• Batch size: 32

Transfer Learning Evaluation

Transfer learning models (MobileNetV2 and VGG16) were evaluated on a smaller subset of the dataset. The results are as follows:

• MobileNetV2:

  • Test Accuracy: 0.5699999928474426
  • Precision: 0.5869781144781144
  • Recall: 0.57
  • F1-Score: 0.5679252766473905
  • ROC-AUC: 0.8973518931972926

• VGG16:

  • Test Accuracy: 0.46000000834465027
  • Precision: 0.48210802742149483
  • Recall: 0.46
  • F1-Score: 0.45696542882981744
  • ROC-AUC: 0.8721693513561938

Evaluation

Confusion matrices are generated to visualize the model's performance. The matrices show the true vs. predicted classes for both MobileNetV2 and VGG16.

Conclusion

This project demonstrates the effectiveness of various techniques in optimizing image classification models. Transfer learning with MobileNetV2 and VGG16, combined with hyperparameter tuning and K-Fold Cross-Validation, provides insights into improving model performance on the CIFAR-10 dataset.