Dog Thermal CNN Image Classifier

Overview

Features

• Transfer Learning: Uses MobileNetV2 for efficient training
• Fast Training: Only 10 epochs with early stopping
• Data Augmentation: Improves model generalization
• Ready for Deployment: Saves trained model in H5 format
• Inference Script: Easy-to-use prediction script

Model Architecture

• Base: MobileNetV2 (pre-trained on ImageNet)
• Custom layers: Global Average Pooling + Dense layers
• Optimized for speed and accuracy

Installation

1. Install dependencies:

Usage

Training the Model

• Load and preprocess the dataset
• Train the model with data augmentation
• Save the best model to dog_thermal_classifier.h5
• Generate training plots and metrics
• Create model info JSON file

Making Predictions

Output Files

• dog_thermal_classifier.h5 - Trained model
• model_info.json - Model metadata and performance metrics
• training_history.json - Training history data
• training_plots.png - Accuracy and loss plots

Model Configuration

• Image Size: 224x224 pixels
• Batch Size: 32
• Epochs: 10 (with early stopping)
• Optimizer: Adam
• Learning Rate: 0.001

Performance Optimizations

1. MobileNetV2: Lightweight architecture for fast inference
2. Reduced Epochs: Only 10 epochs to minimize training time
3. Early Stopping: Stops training when validation loss plateaus
4. Frozen Base Layers: Speeds up training by freezing pre-trained weights
5. Efficient Data Loading: Uses ImageDataGenerator for memory efficiency

Dataset Structure

c:\CNN\
├── train/          # Training images
├── valid/          # Validation images
├── test/           # Test images
└── ...

Ready for Deployment

• Loaded with TensorFlow/Keras
• Converted to TensorFlow Lite for mobile deployment
• Deployed to cloud services (AWS, GCP, Azure)
• Used in web applications with TensorFlow.js

Notes

• The model uses transfer learning for faster training
• Data augmentation helps prevent overfitting
• Early stopping prevents unnecessary training time
• The model is optimized for both speed and accuracy