Sports Celebrity Image Classification: Data Cleaning

This project focuses on classifying images of sports celebrities (and other notable figures) using a machine learning pipeline. The initial step involves data cleaning and preprocessing, where raw images are processed to detect and crop faces with at least two visible eyes. Features are then extracted using wavelet transforms to prepare the data for training a classifier.

Project Overview

• Objective: Classify images of celebrities by preprocessing them to detect and crop facial regions, then extracting features for machine learning.
• Dataset: Custom dataset of celebrity images stored in the dataset/ directory.
• Preprocessing: Face detection using Haar cascades, eye detection, cropping, and wavelet transform for feature extraction.
• Model: SVM with RBF kernel (tuned via GridSearchCV) as the primary classifier.
• Evaluation: Confusion matrix, precision, recall, and F1-score.

Dataset

The dataset consists of images organized in subdirectories under dataset/, each named after a celebrity (e.g., lionel_messi, maria_sharapova, etc.). The preprocessing pipeline crops facial regions and saves them in dataset/cropped/ for model training. Example celebrities include:

• Lionel Messi
• Maria Sharapova
• Roger Federer
• Serena Williams
• Virat Kohli

Sample image preprocessing:

• Input: Raw image (e.g., sharapova1.jpg - 555x700x3).
• Output: Cropped facial region (if two eyes are detected).

Prerequisites

To run this project, you need the following Python libraries:

• numpy
• opencv-python (cv2)
• matplotlib
• pywavelets (PyWavelets)
• scikit-learn
• seaborn (for visualization)

Install the dependencies using pip:

pip install numpy opencv-python matplotlib pywavelets scikit-learn seaborn

Additionally, download the Haar cascade files for face and eye detection from the OpenCV GitHub repository and place them in opencv/haarcascades/:

• haarcascade_frontalface_default.xml
• haarcascade_eye.xml

How to Run

1. Clone the Repository:

   git clone https://github.com/Monish-Nallagondalla/Image-Classification.git
   cd Image-Classification

2. Prepare the Dataset:

   • Place raw celebrity images in subdirectories under dataset/ (e.g., dataset/lionel_messi/).
   • Ensure the opencv/haarcascades/ directory contains the Haar cascade XML files.

3. Run the Code:

   • Open the .ipynb files in Jupyter Notebook:

     jupyter notebook

   • Execute the cells in the following notebooks:
     1. Data cleaning and preprocessing (face detection, cropping, wavelet transform).
     2. Model training and evaluation (SVM, GridSearchCV).

4. Output:

   • Cropped images saved in dataset/cropped/.
   • Visualizations of raw and cropped images.
   • Model performance metrics (e.g., accuracy, confusion matrix).

Code Breakdown

1. Data Cleaning and Preprocessing

• Face Detection: Uses haarcascade_frontalface_default.xml to detect faces in grayscale images.
• Eye Detection: Uses haarcascade_eye.xml to ensure at least two eyes are visible.
• Cropping: Crops the facial region if two eyes are detected, discarding obstructed or unclear images.
• Wavelet Transform: Applies a 5-level Daubechies wavelet transform (db1) to extract edge features.

Key function:

def get_cropped_image_if_2_eyes(image_path):
    img = cv2.imread(image_path)
    gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
    faces = face_cascade.detectMultiScale(gray, 1.3, 5)
    for (x, y, w, h) in faces:
        roi_gray = gray[y:y+h, x:x+w]
        roi_color = img[y:y+h, x:x+w]
        eyes = eye_cascade.detectMultiScale(roi_gray)
        if len(eyes) >= 2:
            return roi_color

• Output: Cropped images saved in dataset/cropped/<celebrity_name>/ (e.g., messi1.png).

2. Feature Extraction

• Raw Images: Resized to 32x32 pixels (3072 features: 32323).
• Wavelet Transform: Resized to 32x32 pixels (1024 features: 32*32).
• Combined Features: Stacked vertically to form a 4096-dimensional vector per image.

3. Model Training

• Pipeline: StandardScaler + SVM (RBF kernel, C=10).
• GridSearchCV: Evaluates SVM, Random Forest, and Logistic Regression with various parameters.
• Best Model: SVM with RBF kernel (accuracy: ~75%).

4. Evaluation

• Confusion Matrix:

[[21  1  1  0  3  0]
 [ 0 12  2  0  2  2]
 [ 1  2 30  0  1  1]
 [ 1  0  0 24  2  1]
 [ 2  0  2  1 19  2]
 [ 2  3  4  0  0  4]]

• Classification Report:

              precision    recall  f1-score   support
     0 (Elon Musk)     0.78      0.81      0.79        26
     1 (Kanye West)    0.67      0.67      0.67        18
     2 (Megan Fox)     0.77      0.86      0.81        35
     3 (Modi)          0.96      0.86      0.91        28
     4 (Putin)         0.70      0.73      0.72        26
     5 (Travis Scott)  0.40      0.31      0.35        13
    accuracy                           0.75       146

Results

• Preprocessing: Successfully cropped images with two visible eyes, discarding obstructed ones.
• Model Performance: Achieved ~75% accuracy with SVM (RBF kernel). Modi (class 3) has the highest precision (0.96), while Travis Scott (class 5) has the lowest recall (0.31).
• Visualizations: Heatmap of confusion matrix highlights classification performance.

Future Improvements

• Expand the dataset with more sports celebrities and images.
• Implement data augmentation (e.g., rotation, flipping) to increase robustness.
• Experiment with deep learning models (e.g., CNNs) for better feature extraction.
• Fine-tune wavelet transform parameters or explore other feature extraction methods.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

• Debjyoti Paul for his invaluable insights and support.
• OpenCV documentation: Face Detection Tutorial.
• Built with Python, OpenCV, PyWavelets, scikit-learn, and matplotlib.

Repository

GitHub Repository

### Instructions
1. Copy the entire code block above.
2. Open or create a `README.md` file in your GitHub repository (`https://github.com/Monish-Nallagondalla/Image-Classification`).
3. Paste the code into the file.
4. Ensure the dataset structure (`dataset/` and `opencv/haarcascades/`) matches the paths in the code.
5. Save and commit the file to your repository.