A machine learning project focused on detecting and classifying car damage using deep learning models. This project compares two different architectures for car damage detection, utilizing a dataset of car images from Kaggle. The models were trained and evaluated using TensorFlow and Keras in a Jupyter Notebook environment.
The goal of this project is to accurately detect and classify car damage using a machine learning-based approach. We aim to compare the performance of two different models:
- Model 1: MobileNetV2 (Pre-trained)
- Model 2: Custom CNN (Convolutional Neural Network)
Both models are assessed based on their accuracy, performance, and resource efficiency.
- Source: Kaggle Car Damage Dataset
- Size: 1,920 images of car damages
- Classes: Damage / No Damage
├── data/ # Dataset and preprocessed files
├── models/ # Saved model files
├── notebooks/ # Jupyter notebooks
├── images/ # Plots, graphs, and result images
├── README.md # Project documentation
└── requirements.txt # Required Python libraries- Language: Python
- Libraries: TensorFlow, Keras, NumPy, Pandas, Matplotlib, Scikit-learn, OS
- Jupyter Notebook: For interactive code and visualizations
- Transfer Learning: MobileNetV2 used as a base model in Model 1
-
Clone the Repository:
git clone <repo-url> cd car-damage-detection
-
Install Dependencies:
pip install -r requirements.txt
-
Download Dataset: Place the dataset in the
data/directory. You can find the dataset on Kaggle. -
Run the Jupyter Notebook: Open the notebook and run the cells to preprocess data, train the models, and visualize results.
- Pre-trained MobileNetV2 architecture
- Used transfer learning to fine-tune the model
- Key steps: Loading base model, attaching custom head, training on car damage dataset
- Convolutional Neural Network (CNN) built from scratch
- Key steps: Convolutional layers for feature extraction, dense layers for classification, with dropout regularization
- Architecture: Conv -> ReLU -> MaxPool -> Dropout -> Fully Connected -> Softmax
- Preprocessing: Images were preprocessed (resize, normalize, label encoding).
- Data Augmentation: Used
ImageDataGeneratorto perform augmentation (rotation, zoom, etc.). - Model Training: Both models were trained and validated on the Kaggle dataset.
- Results Visualization: Accuracy and loss plots were generated for training and validation data.
| Metric | MobileNetV2 |
|---|---|
| Accuracy | 78.7% |
| Parameters | 2,422,210 |
- MobileNetV2 showed better accuracy with a faster inference time due to transfer learning.
- Custom CNN required more time to train but offered more flexibility in terms of tuning the architecture.
Training and testing accuracy/loss graphs for both models:
)
+----------------------------------------------------------+
| Start |
+----------------------------------------------------------+
|
v
+----------------------------------------------------------+
| Import Necessary Libraries |
| Tools: TensorFlow, Keras, NumPy, Pandas |
| Why: TensorFlow/Keras for model building, |
| NumPy for array operations, Pandas for dataframes |
+----------------------------------------------------------+
|
v
+----------------------------------------------------------+
| Define Image Paths and Label Categories |
| Tools: OS (for file handling), Pandas |
| Why: OS helps manage the directory structure, |
| Pandas assists with organizing the data labels |
+----------------------------------------------------------+
|
v
+----------------------------------------------------------+
| Check if Paths and Categories Exist |
| Tools: OS |
| Why: To verify if the correct directories and files are |
| available before loading data |
+----------------------------------------------------------+
|
v
+----------------------------------------------------------+
| Load Images from Directory |
| Tools: Keras (image_dataset_from_directory) |
| Why: Efficiently loads and labels images while scaling |
| them into a usable format |
+----------------------------------------------------------+
|
v
+----------------------------------------------------------+
| Preprocess Images |
| Tools: TensorFlow (tf.image), Keras |
| Why: Resize images to 224x224 (for consistency) and |
| normalize pixel values for faster model convergence|
+----------------------------------------------------------+
| Resize Images (224x224) |
| Normalize Pixel Values (0 to 1 scale) |
+----------------------------------------------------------+
|
v
+----------------------------------------------------------+
| One-Hot Encode Labels |
| Tools: TensorFlow (tf.keras.utils.to_categorical) |
| Why: Converts categorical labels into a binary vector |
| form required for classification tasks |
+----------------------------------------------------------+
|
v
+----------------------------------------------------------+
| Split Data into Training and Testing Sets |
| Tools: scikit-learn (train_test_split) |
| Why: To create a reliable 80-20 training-test split, |
| ensuring the model doesn’t learn the test data |
+----------------------------------------------------------+
| 80% Training, 20% Testing |
| Shuffle and ensure no data leakage |
+----------------------------------------------------------+
|
v
+----------------------------------------------------------+
| Apply Data Augmentation (Optional) |
| Tools: TensorFlow (ImageDataGenerator) |
| Why: Enhances training data by applying random |
| transformations (flips, rotations, zooms) to avoid|
| overfitting |
+----------------------------------------------------------+
|
v
+----------------------------------------------------------+
| Load Pre-Trained MobileNetV2 Model |
| Tools: Keras (keras.applications.MobileNetV2) |
| Why: MobileNetV2 is a lightweight pre-trained model that|
| is efficient for image classification tasks |
+----------------------------------------------------------+
| (exclude top layers - include_top=False) |
+----------------------------------------------------------+
|
v
+----------------------------------------------------------+
| Freeze Base Model Layers |
| Tools: Keras (model.trainable = False) |
| Why: Prevents pre-trained layers from being modified, |
| ensuring their learned weights are preserved |
+----------------------------------------------------------+
|
v
+----------------------------------------------------------+
| Add Custom Dense Layers |
| Tools: Keras (Sequential, Dense, Flatten, Softmax) |
| Why: To adapt the model for your specific classification|
| task by adding custom layers on top of MobileNetV2 |
+----------------------------------------------------------+
| Flatten layer + Dense Layer + Softmax Output |
+----------------------------------------------------------+
|
v
+----------------------------------------------------------+
| Compile the Model |
| Tools: Keras (model.compile) |
| Why: Set the loss function, optimizer, and metrics to |
| guide how the model trains |
+----------------------------------------------------------+
| Loss: categorical_crossentropy |
| Optimizer: Adam |
| Metrics: accuracy |
+----------------------------------------------------------+
|
v
+----------------------------------------------------------+
| Train the Model |
| Tools: Keras (model.fit) |
| Why: Trains the model using the training data with |
| specified batch size, epochs, and augmentation |
+----------------------------------------------------------+
| Batch Size: 32 |
| Number of Epochs: 20 |
| Save training history for later analysis |
+----------------------------------------------------------+
|
v
+----------------------------------------------------------+
| Evaluate the Model on Test Data |
| Tools: Keras (model.evaluate) |
| Why: Evaluates model performance on the test dataset |
| to calculate accuracy, precision, recall, etc. |
+----------------------------------------------------------+
|
v
+----------------------------------------------------------+
| Save the Trained Model |
| Tools: Keras (model.save) |
| Why: Save the model for future use or deployment |
+----------------------------------------------------------+
| Use model.save('model_name.h5') |
+----------------------------------------------------------+
|
v
+----------------------------------------------------------+
| Plot Training Results |
| Tools: Matplotlib, Seaborn |
| Why: Visualize training and validation loss and accuracy|
| to monitor the learning process |
+----------------------------------------------------------+
| Plot loss and accuracy for training vs validation sets |
+----------------------------------------------------------+
|
v
+----------------------------------------------------------+
| Make Predictions on New Data |
| Tools: Keras (model.predict), NumPy |
| Why: Make predictions on new/unseen data to validate |
| the model’s ability to generalize |
+----------------------------------------------------------+
|
v
+----------------------------------------------------------+
| Generate Confusion Matrix and Classification Report |
| Tools: Scikit-learn (confusion_matrix, classification_report) |
| Why: To better understand the model's performance with |
| precision, recall, F1-score, and class-wise errors |
+----------------------------------------------------------+
|
v
+----------------------------------------------------------+
| End |
+----------------------------------------------------------+- Car Damage Detection and Classification
- Car Damage Assessment Based on VGG Models
- A Very Deep Transfer Learning Model for Vehicle
- Further fine-tuning of the models for higher accuracy
- Exploring additional architectures like EfficientNet
- Expanding the dataset to include more diverse types of car damage
This project provides a comparative study between MobileNetV2 and a Custom CNN for car damage detection. While both models performed well, the choice of model depends on application requirements like accuracy, speed, and available computational resources.