Image Classification using MobileNetV2

Machine Learning Model for Classifying Images into Four Categories

---

1. Project Overview

This project implements an image classification system using deep learning. The goal is to train a model that classifies images into four categories based on the folder names in the dataset.

Task: You can complete the following task using any programming language and ML framework as you wish.

1. Create an image preprocessing pipeline that prepares the images and potentially new images to be passed through an ML model that you will design in the second step.

2. Design, write, and train an ML model that will perform the classification of the provided image data. The classes required correspond to the folder name of the provided dataset. You should also include either in comments or in a separate file your reasoning for choosing said ML model and document your training approach.

3. The trained model should be exported. A separate program be written that can accept an image not previously included in the dataset and classification be performed on it using the trained model.

For this task, you will be using the images provided in the “images” folder of the provided dataset. The datasets to be used for the task are available here: https://drive.google.com/drive/folders/1UtgQWE3AO0GLvom-osrJVQ7M6y83MajF?usp=sharing

My structure preparation for this task included:

1. Preprocessing images to prepare them for deep learning.
2. Training a MobileNetV2-based neural network for classification.
3. Deploying the model to recognise unseen images based on trained categories.

Why MobileNetV2?
The decision to use in this project MobileNetV2 was made regarding the following reasons:
✅ Pre-trained on ImageNet – Works well even with a small dataset.
✅ Lightweight and Efficient – Ideal for real-time classification.
✅ Good Generalization – Transfer learning helps adapt to new datasets quickly.

---

2. Dataset & Preprocessing

📂 Dataset Structure

The dataset consists of images structured into folders, where each folder name represents the class label:

Images/
│── bicycles/
│── car/
│── deer/
│── mountains/

Preprocessing Steps

The following steps were added to improve model accuracy:
✅ Removing corrupted images (some images were found to be truncated).
✅ Resizing images to 224x224 (required by MobileNetV2).
✅ Scaling pixel values to [0,1] range for better training.
✅ Splitting dataset into training (80%) and validation (20%).

Issues Encountered & Fixes:
1️⃣ Corrupted images (OSError: image file is truncated) → Automatically detected and removed.
2️⃣ .DS_Store files appearing in datasets → Ignored in preprocessing.

---

3. Model Selection: Why MobileNetV2?

Why not a Custom CNN?

• Custom CNNs take longer to train and require more data to generalise well.
• MobileNetV2 is pre-trained on ImageNet, making it an ideal transfer learning approach.

Why MobileNetV2?

✅ Pre-trained on ImageNet – Works well even with limited data.
✅ Highly Efficient – Uses depthwise separable convolutions to be lightweight and fast.
✅ Great Generalization – Transfer learning allows quick adaptation to new datasets.

Model Architecture:

• MobileNetV2 base model (pre-trained on ImageNet).
• Global Average Pooling layer to reduce feature dimensions.
• Fully connected (Dense) layers with ReLU and Softmax activation.
• Dropout layers to prevent overfitting.

---

4. Training Approach & Hyperparameters

Hyperparameters Used:

• Optimiser: Adam (lr=0.001 initially, dynamically reduced)
• Loss Function: categorical_crossentropy (for multi-class classification)
• Batch Size: 32
• Epochs: 20 (Early stopping applied)
• Learning Rate Reduction: ReduceLROnPlateau (auto-adjusts learning rate)
• Regularisation: Dropout layers to prevent overfitting

Training Enhancements:

✅ EarlyStopping → Stops training if validation loss stops improving.
✅ ReduceLROnPlateau → Lowers learning rate if model performance plateaus.

---

5. Results & Analysis

Final Model Performance:

Metric	Training	Validation
Accuracy	99.9%	93.3%
Loss	0.0449	0.2496

Observations:

✔ Good Generalisation – Validation accuracy is close to training accuracy.
✔ Overfitting Controlled – ReduceLROnPlateau helped avoid overfitting.
✔ Fast Convergence – High accuracy achieved in just 7 epochs.

Potential Improvements:
1️⃣ More data augmentation techniques could be applied to further improve generalization.
2️⃣ Training for a few more epochs with gradual learning rate decay could refine results.

---

6. Issues that have been fixed during progress

1. Image File Corruption Errors

• Error: OSError: image file is truncated (51 bytes not processed)
• Fix: A function was implemented to scan for and remove corrupted images before training.

2. Unexpected .DS_Store Files

• Error: macOS automatically creates .DS_Store files in directories.
• Fix: The code was updated to ignore non-image files when processing data.

3. Model Architecture Warning

• Error: UserWarning: Do not pass an input_shape argument to a layer.
• Fix: The model was modified to use an Input layer explicitly instead of passing input_shape in layers.

4. Classification Predictions Returning Numerical Labels

• Error: Predicted class: 0 instead of cars, deer, etc.
• Fix: The final prediction output was mapped to class labels.

---

7. How to Train & Use the Model

Step 1️⃣: Train the Model

Run the following command:

python image_prepare.py

• Select 'train' when prompted.
• The model will train and save as trained_model.keras.

Step 2️⃣: Classify an Image

Run the following command:

python image_prepare.py

• Select 'recognise' and enter the image path.
• Example Output:

  Predicted class: Car (99.5% confidence)
  

Step 3️⃣: Exit

After finishing with either training the model or classifying an Image or both, you have an option to exit.

• Select 'exit'.

---

8. Installation & Setup

1️⃣ Install Dependencies

pip install tensorflow keras numpy pillow

2️⃣ Set Up a Virtual Environment (Recommended)

To avoid dependency issues, create a new virtual environment:

python -m venv myenv
source myenv/bin/activate  # Mac/Linux
myenv\Scripts�ctivate  # Windows

Then, reinstall the dependencies:

pip install tensorflow keras numpy pillow

---

Final Notes

✅ The model achieves 93%+ validation accuracy.
✅ For best performance, run this in a dedicated virtual environment (venv).

---

9. Future Improvements

Fine-tuning MobileNetV2 layers to improve performance.
Experimenting with data augmentation to improve generalization.
Using a larger dataset for better accuracy.