Scene Classification Model

A deep learning model for classifying different types of scenes using TensorFlow and ResNet50V2. The model is designed to classify scenes into 6 different categories.

Project Structure

.
├── data/
│   ├── train/
│   │   └── [image files]
│   ├── train.csv
│   └── test.csv
├── models/
│   └── version_YYYYMMDD_HHMMSS/
│       ├── model.h5
│       ├── training_history.csv
│       └── model_summary.txt
├── train.py
├── predict.py
└── README.md

Requirements

• Python 3.8+
• TensorFlow 2.x
• OpenCV
• NumPy
• Pandas
• Matplotlib
• scikit-learn

Install the required packages using:

pip install tensorflow opencv-python numpy pandas matplotlib scikit-learn

Dataset Structure

Training Data

• train.csv: Contains two columns:
  • image_name: Name of the image file
  • label: Integer label (0-5) representing the scene category

Test Data

• test.csv: Contains one column:
  • image_name: Name of the image file

Image Directory

• All images should be placed in the data/train/ directory
• The model expects RGB images (any common format like jpg, png)
• Images will be automatically resized to 224x224 pixels during processing

Training the Model

The training script (train.py) handles:

• Data loading and preprocessing
• Model creation using ResNet50V2
• Training with data augmentation
• Model saving with versioning

To train the model:

python train.py

Training configurations (in train.py):

• Image size: 224x224
• Batch size: 32
• Epochs: 10
• Training/Validation split: 80/20
• Data augmentation:
  • Rotation range: 20°
  • Width/Height shift: 20%
  • Horizontal flip

The trained model will be saved in models/version_YYYYMMDD_HHMMSS/ with:

• Model file (model.h5)
• Training history (training_history.csv)
• Model architecture summary (model_summary.txt)

Making Predictions

The prediction script (predict.py) offers two modes:

1. Single Image Prediction
2. Full Test Set Prediction

To use the predictor:

python predict.py

Single Image Prediction

• Loads and displays the image
• Shows prediction with confidence score
• Visualizes results using matplotlib

Test Set Prediction

• Processes all images in test.csv
• Saves predictions to a CSV file with:
  • Image name
  • Predicted class
  • Confidence score

The script automatically uses the latest trained model, but you can modify the code to use a specific version.

Model Architecture

The model uses ResNet50V2 as the base architecture with additional layers:

• ResNet50V2 (pretrained on ImageNet)
• Global Average Pooling
• Dense layer (256 units, ReLU)
• Dropout (0.5)
• Output layer (6 units, Softmax)

Training parameters:

• Optimizer: Adam
• Loss: Sparse Categorical Crossentropy
• Metric: Accuracy

Error Handling

Both scripts include comprehensive error handling:

• File not found errors
• Image processing errors
• Model loading/saving errors
• Training/prediction errors

All errors are logged with timestamps for easy debugging.

Best Practices

1. Data Organization:

   • Keep all images in the data/train/ directory
   • Ensure image names match those in CSV files
   • Maintain consistent image formats

2. Model Versioning:

   • Each training run creates a new versioned directory
   • Keep track of different model versions for comparison
   • Don't delete old versions until you're sure they're not needed

3. Resource Management:

   • Large datasets might require batch processing
   • Consider GPU availability for training
   • Monitor memory usage with large test sets

Troubleshooting

Common issues and solutions:

1. Memory Errors:

   • Reduce batch size in train.py
   • Process test set in smaller chunks
   • Close other memory-intensive applications

2. Image Loading Errors:

   • Check image file permissions
   • Verify image format compatibility
   • Ensure correct file paths

3. Model Loading Errors:

   • Verify model directory exists
   • Check for corrupted model files
   • Ensure compatible TensorFlow versions

Contributing

Feel free to submit issues and enhancement requests!