Image Classification using Intel Image Dataset

Project Overview

Developed a multi-class image classification system using the Intel Image Classification dataset, focusing on data exploration, custom CNN architectures, hyperparameter tuning, and transfer learning with MobileNet. The project showcases comparative evaluation to identify the most effective model.

---

Table of Contents

1. Data Exploration & Preprocessing
2. Model Development
   • Simple Classifier
   • Complex Classifier with Keras-Tuner
   • Custom ConvNet Architecture
   • Pre-trained Model - MobileNet
3. Comparative Evaluation
4. Results
5. Conclusion
6. How to Run the Code
7. References

---

1. Data Exploration & Preprocessing

• Analyzed six image classes: Street, Sea, Mountain, Glacier, Forest, and Buildings to ensure balanced data distribution between training and test datasets.
• Verified class proportions between datasets to ensure consistency and avoid model overfitting.
• Loaded datasets using TensorFlow's image_dataset_from_directory function, preparing them for training, validation, and testing.

---

2. Model Development

2.1 Simple Classifier

• Developed a basic neural network with:
  • Flatten layer and appropriate output layer based on the number of classes.
  • Implemented early stopping using 20% of the training set for validation.
• Reported initial test accuracy and identified areas for improvement.

2.2 Complex Classifier with Keras-Tuner

• Tuned hyperparameters with Keras-Tuner, experimenting with:
  • Number and sizes of hidden layers.
  • Dropout rates for regularization.
  • Learning rates to ensure smooth convergence.

2.3 ConvNet Architecture Implementation

• Built a custom CNN architecture with:
  • Four Conv2D layers followed by MaxPooling2D for feature extraction.
  • Optimized hyperparameters like filter sizes, kernel dimensions, and pool sizes through experimentation.
• Achieved 54.67% test accuracy with this custom architecture.

2.4 Pre-trained Model - MobileNet

• Used MobileNet, pre-trained on ImageNet, to perform transfer learning:
  • Frozen MobileNet weights during training and added a custom classification layer.
  • Trained the model on the Intel dataset and achieved 91.07% accuracy on the test set.

---

3. Comparative Evaluation

• Conducted comparative analysis between the custom CNN model and the MobileNet model.
• Visualized results using a confusion matrix to identify misclassifications.
• Noted MobileNet’s superior performance with:
  • MobileNet Accuracy: 91.07%
  • Custom CNN Accuracy: 54.67%
• Displayed sample images with the most common classification errors for further analysis.

---

4. Results

• Best Model: MobileNet with 91.07% accuracy on the test set.
• Common Misclassifications: Errors primarily involved distinguishing between similar categories, such as mountain vs. glacier and street vs. buildings.

---

5. Conclusion

• The MobileNet model demonstrated significant improvement over the custom CNN, emphasizing the effectiveness of transfer learning.
• The project highlights the importance of hyperparameter tuning, model selection, and data consistency for achieving high accuracy in multi-class image classification tasks.

---

6. How to Run the Code

1. Clone the repository:

   git clone https://github.com/yourusername/intel-image-classification.git
   cd intel-image-classification

2. Install dependencies:

   pip install -r requirements.txt

7. Reference

• Intel Image Classification Dataset: Intel Open Data
• TensorFlow Documentation: TensorFlow
• MobileNet Paper: Howard et al., MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications