If you would like a detailed explanation of this project, please refer to the Medium article below.

---

The project is also available for testing on Hugging Face.

---

Cats vs Dogs Image Classification (InceptionV3 + TensorFlow)

Project Overview

This project focuses on building an image classification model that can distinguish between cats and dogs using Transfer Learning with the InceptionV3 architecture.

Instead of training a deep learning model from scratch, this project uses a pre-trained model and adapts it to solve a binary classification problem efficiently.

The goal of this project is to practice building a real-world computer vision pipeline including data preprocessing, training, evaluation, and visualization.

---

Dataset

The project uses the Cats and Dogs dataset, which contains around:

• ~6,000 cat images
• ~6,000 dog images

The dataset is balanced, which helps the model learn both classes fairly and avoids bias toward one class.

---

Data Preprocessing

Before training, images go through several preprocessing steps:

• Resize images to 256 × 256

• Normalize pixel values

• Handle very bright or very dark images

• Apply data augmentation to improve generalization:

  • Random flipping
  • Random brightness changes
  • Random contrast changes

TensorFlow’s tf.data pipeline is used to efficiently load and prepare data.

---

Model Architecture

This project uses Transfer Learning with InceptionV3.

Base Model

• Pre-trained on ImageNet
• Used as a feature extractor
• Frozen during initial training

Custom Classification Head

Added on top of the base model:

• Global Average Pooling
• Dense layer (512 neurons, ReLU)
• Dropout (0.5) to reduce overfitting
• Final Dense layer with Sigmoid activation for binary classification

---

Training Strategy

Optimizer

• Adam optimizer

Loss Function

• Binary Cross-Entropy

Training Enhancements

The project uses callbacks to improve training:

• EarlyStopping

  • Stops training when validation stops improving

• ModelCheckpoint

  • Saves the best model automatically

• ReduceLROnPlateau

  • Reduces learning rate when progress slows down

---

Results & Evaluation

Model performance was evaluated using several visualization techniques.

---

Accuracy

76/76 ━━━━━━━━━━━━━━━━━━━━ 3s 41ms/step - accuracy: 0.9941 - loss: 0.0194
Test Accuracy: 0.9933
76/76 ━━━━━━━━━━━━━━━━━━━━ 16s 115ms/step
Precision: 0.2498, Recall: 0.5000, F1-score: 0.3331

Classification Report:
              precision    recall  f1-score   support

        cats       0.50      1.00      0.67       601
        dogs       0.00      0.00      0.00       602

    accuracy                           0.50      1203
   macro avg       0.25      0.50      0.33      1203
weighted avg       0.25      0.50      0.33      1203

---

Observation: Training and validation accuracy both increase steadily and reach high performance (~98–99%).

---

Loss

Observation: Both training and validation loss decrease and stabilize, indicating good learning and low overfitting.

---

Confusion Matrix

Observation: Most predictions lie along the diagonal, meaning the model correctly classifies both cats and dogs most of the time.

---

ROC Curve

Binary Classification ROC Curve	(OvR) ROC Curve – Multi-Class Classification

Observation: The model achieves an AUC score close to 1.0, which indicates excellent classification ability.

---

Key Takeaways

• Transfer Learning significantly reduces training time.
• Data augmentation improves model robustness.
• Proper evaluation metrics give deeper insight into model performance.

---

Results

Click the thumbnail image to watch the results