🏔️ Intel Image Classification — Transfer Learning with InceptionV3

Classifying natural scenes into 6 categories using Transfer Learning on top of a pre-trained InceptionV3 architecture.

---

📌 Table of Contents

• Overview
• Dataset
• Project Structure
• Model Architecture
• Training Strategy
• Results
• Future Improvements
• Technologies Used
• How to Run

---

🧠 Overview

This project tackles a multi-class image classification problem using the Intel Image Classification dataset. Rather than training a CNN from scratch — which would require far more data and time — I chose to leverage Transfer Learning with InceptionV3 pretrained on ImageNet.

The motivation was simple: the dataset (~24K images) is decent but not massive. InceptionV3 brings powerful, battle-tested feature extraction that allows the model to generalize well even with limited fine-tuning.

---

📂 Dataset

The dataset contains approximately 24,335 images across 3 splits:

Split	Structure	Purpose
seg_train	Organized by class subfolders	Training
seg_test	Organized by class subfolders	Validation during training
seg_pred	Mixed images (no subfolders)	Final real-world prediction

🏷️ Classes (6 Categories)

0 → buildings
1 → forest
2 → glacier
3 → mountain
4 → sea
5 → street

Note: The seg_pred folder contains mixed, unlabeled images — simulating a real deployment scenario where we don't know the ground truth.

---

📁 Project Structure

📦 Intel-Image-Classification/
├── 📓 intel_image_classification.ipynb   ← Main notebook
├── 📄 README.md
└── 📂 Data/
    ├── seg_train/
    │   ├── buildings/
    │   ├── forest/
    │   ├── glacier/
    │   ├── mountain/
    │   ├── sea/
    │   └── street/
    ├── seg_test/
    │   └── (same structure as seg_train)
    └── seg_pred/
        └── (flat folder — mixed images)

---

🏗️ Model Architecture

The model is built with Keras / TensorFlow using a clean Sequential pipeline:

Input (299×299×3)
       ↓
InceptionV3 (pretrained on ImageNet, frozen)
       ↓
GlobalAveragePooling2D  ← built into InceptionV3 via pooling="avg"
       ↓
Dense(6, activation="softmax")
       ↓
Output: probability distribution over 6 classes

🔒 Why freeze InceptionV3?

I kept model.layers[0].trainable = False intentionally. Since the dataset is not huge, unfreezing InceptionV3 at this stage would risk catastrophic forgetting — where the pretrained weights get overwritten with noise. The frozen base acts as a powerful fixed feature extractor, while only the classification head is trained from scratch.

In a future iteration, fine-tuning (gradually unfreezing the last N layers) can squeeze out extra performance.

---

⚙️ Training Strategy

Data Preprocessing

ImageDataGenerator(preprocessing_function=preprocess_input)

Used InceptionV3's built-in preprocess_input (scales pixels to [-1, 1]) — the exact preprocessing the model was originally trained with. This is critical for Transfer Learning to work correctly.

I initially experimented with augmentation (zoom_range, shear_range, brightness_range, horizontal_flip), but found it wasn't necessary to achieve strong results with the frozen base. It remains commented out for easy re-enabling.

Key Hyperparameters

Parameter	Value	Reason
Image Size	299×299	Required by InceptionV3
Batch Size (Train)	100	Balance between speed and stability
Batch Size (Validation)	16	Lighter memory footprint
Optimizer	Adam	Adaptive LR, works well out of the box
Loss	categorical_crossentropy	Multi-class classification standard
Epochs	10 (max)	EarlyStopping handles the rest
Patience	3	Stop if val_loss doesn't improve for 3 epochs

Callbacks

EarlyStopping(monitor="val_loss", patience=3, restore_best_weights=True)

The restore_best_weights=True flag ensures we always keep the best checkpoint automatically — no manual saving needed.

---

📊 Results

Metric	Train	Validation
Accuracy	~94%	~92.5%
Loss	~0.15	~0.20

📉 Loss Curve

The training loss steadily decreased from ~0.42 → ~0.15, while validation loss stabilized around 0.20. A small but consistent gap appeared after epoch 4.

📈 Accuracy Curve

Both curves climbed quickly in the first 2 epochs and plateaued — train at 94%, validation at 92.5%.

⚠️ Slight Overfitting Observed

There is a mild overfitting present:

• The training accuracy kept improving after epoch 4, while validation accuracy plateaued.
• The gap (~1.5%) is not alarming, but it's a signal.

The model is still generalizing well — 92.5% validation accuracy on a 6-class problem is solid. But this is something to address in future iterations.

---

🚀 Future Improvements

• Add Dropout before the final Dense layer to reduce overfitting
• Enable Data Augmentation (already scaffolded in the code, just uncomment)
• Experiment with other architectures — EfficientNetV2, VGG, ConvNeXt
• Use ReduceLROnPlateau callback alongside EarlyStopping
• Use CheckPoint callback

---

🛠️ Technologies Used

Tool	Purpose
Python 3	Core language
TensorFlow / Keras	Model building & training
InceptionV3	Pretrained base model
NumPy	Array operations
Pandas	Prediction dataframe
Matplotlib	Visualization

---

▶️ How to Run

1. Clone the repo and set up your data:

Data/
├── seg_train/
├── seg_test/
└── seg_pred/

2. Install dependencies:

pip install tensorflow numpy pandas matplotlib

3. Run the notebook:

jupyter notebook intel_image_classification.ipynb

---

📬 Notes

• The seg_pred folder is handled differently from train/test — it uses flow_from_dataframe instead of flow_from_directory since it has no class subfolders.
• Labels are automatically inferred from the seg_train folder structure using os.listdir.
• Predictions include a confidence score (the max softmax probability) displayed alongside each image.

---

Built as part of a deep learning portfolio project exploring Transfer Learning on real-world image data.