Intel Image Classification — CNN & Transfer Learning

Image classification of natural scenes into 6 classes (buildings, forest, glacier, mountain, sea, street) built with TensorFlow / Keras. The project compares a custom CNN baseline against a MobileNetV2 transfer-learning model with two-phase fine-tuning, and exports the final model to three deployment formats: SavedModel, TensorFlow Lite, and TensorFlow.js.

---

Results

Model	Train Acc	Val Acc	Test Acc
Baseline CNN (from scratch)	91.93%	89.20%	87.57%
MobileNetV2 (transfer learning)	95.24%	94.83%	93.38%

Transfer learning with two-phase fine-tuning substantially improves generalization over the from-scratch baseline while producing a smaller, deployment-friendly model.

---

Dataset

• Source: Intel Image Classification — Kaggle
• Size: ~17,000 images (seg_train + seg_test combined)
• Classes (6): buildings, forest, glacier, mountain, sea, street
• Resolution: non-uniform (no fixed resolution in the source) — see EDA in the notebook for the resolution distribution analysis.

The provider's original split is not used. All images are pooled per class and re-split manually with random_state=42 for reproducibility:

Split	Ratio
Train	80%
Validation	10%
Test	10%

---

Approach

1. Baseline — Custom CNN. Four Conv2D + BatchNorm + MaxPooling + Dropout blocks (filters 32 → 64 → 128 → 256), GlobalAveragePooling2D, then a dense softmax head. Trained from scratch as a reference point.

2. Transfer Learning — MobileNetV2 (final model).

• Phase 1 — Feature extraction: ImageNet-pretrained backbone frozen; only the classification head is trained (lr = 1e-3).
• Phase 2 — Fine-tuning: the top third of the backbone is unfrozen and trained with a very small learning rate (lr = 1e-5) so pretrained features adapt without being destroyed. Backbone BatchNorm layers are kept in inference mode for stability.

Training details

• Input size 224×224, batch size 32.
• Preprocessing via mobilenet_v2.preprocess_input (range [-1, 1]).
• Augmentation (training only): rotation, width/height shift, shear, zoom, horizontal flip.
• Callbacks: EarlyStopping, ModelCheckpoint, ReduceLROnPlateau.

---

Repository Structure

.
├── notebook.ipynb          # End-to-end pipeline: EDA → training → eval → export → inference
├── README.md
├── requirements.txt
├── .gitignore
└── models/
    ├── saved_model/        # TensorFlow SavedModel
    ├── tfjs_model/         # TensorFlow.js
    └── tflite/             # TensorFlow Lite (+ label.txt)

The dataset is downloaded at runtime from Kaggle and is not committed to the repo.

---

How to Run

The notebook is designed for Google Colab (T4 GPU).

1. Open notebook.ipynb in Colab and select a GPU runtime.
2. Provide Kaggle credentials when prompted — either via Colab Secrets (KAGGLE_USERNAME, KAGGLE_KEY) or by uploading kaggle.json. Credentials are never hard-coded.
3. Runtime → Run all.

To run locally:

pip install -r requirements.txt
jupyter notebook notebook.ipynb

---

Model Formats & Deployment

Format	File	Use case
SavedModel	models/saved_model/	TensorFlow Serving / production backends
TF-Lite	models/tflite/model.tflite	Mobile & edge devices (dynamic-range quantized)
TensorFlow.js	models/tfjs_model/model.json	In-browser inference

Class labels: models/tflite/label.txt.

---

Tech Stack

Python · TensorFlow / Keras · MobileNetV2 · scikit-learn · NumPy · Matplotlib · Seaborn · TensorFlow.js · TensorFlow Lite

---

Author

Muhammad Fariz Abizar

---

License

Released under the MIT License — see LICENSE.