PixelProof

PixelProof is a lightweight classical computer vision system that distinguishes between a real-world photograph and a photograph of a screen.

Unlike deep learning approaches, this project relies entirely on handcrafted image-processing features and traditional machine learning, making it lightweight, interpretable, and suitable for CPU-only inference.

Streamlit web demo - https://pixelproof.streamlit.app/

Demo Video - https://drive.google.com/file/d/1_JT8Q0YUC8PODpAyZb7TEHQy9_eQYZAy/view?usp=drive_link

---

Features

The system extracts a variety of handcrafted image features including:

• Laplacian Variance (blur/sharpness)
• Edge Density
• Local Binary Patterns (LBP)
• Gray Level Co-occurrence Matrix (GLCM)
• FFT Peak Analysis
• RGB Color Statistics
• HSV Color Statistics
• Brightness
• Contrast
• Shannon Entropy
• Glare Percentage
• Pixel Grid Score
• Hough Line Features

These features are used to train classical machine learning classifiers.

---

Models Evaluated

The following models were compared using 5-fold Stratified Cross Validation:

• Random Forest
• Support Vector Machine (SVM)
• XGBoost

The model with the highest validation accuracy is automatically selected and saved.

---

Dataset Structure

dataset/
│
├── real/
│   ├── image1.jpg
│   ├── image2.jpg
│   └── ...
│
└── screen/
    ├── image1.jpg
    ├── image2.jpg
    └── ...

---

Project Structure

PixelProof/
│
├── dataset/
│   ├── real/
│   └── screen/
│
├── models/
│   └── best_model.pkl
│
├── src/
│   ├── feature_extraction.py
│   ├── train.py
│   ├── predict.py
│   ├── error_analysis.py
│   └── features.csv
│
├── requirements.txt
└── README.md

---

Installation

Clone the repository

git clone <repository-url>
cd PixelProof

Install dependencies

pip install -r requirements.txt

---

Generate Features

Run

cd src

python feature_extraction.py

This extracts handcrafted features from every image and creates

features.csv

---

Train the Model

Run

python train.py

The script

• Performs 5-fold Stratified Cross Validation
• Compares Random Forest, SVM and XGBoost
• Reports Accuracy, Precision, Recall, F1-score and ROC-AUC
• Saves the best model as

models/best_model.pkl

---

Error Analysis

To inspect misclassified images, run

python error_analysis.py

This generates:

• Confusion Matrix
• Classification Report
• List of misclassified images

---

Predict a New Image

Run

python predict.py path/to/image.jpg

Example

python predict.py dataset/real/sample.jpg

Output

Score      : 0.0021

or

Score      : 0.9984

The score represents the probability that the input image is a photograph of a screen.

• Score close to 0 → Real Photo
• Score close to 1 → Screen Photo

---

Performance

Using 5-fold Stratified Cross Validation:

Metric	Score
Accuracy	~96%
Precision	~96%
Recall	~96%
F1 Score	~96%
ROC-AUC	~0.99

---

Inference Time

Measured on a Windows laptop CPU.

Stage	Average Time
Feature Extraction	~250–300 ms
Prediction	~60–90 ms
Total	~320–390 ms

---

Technologies Used

• Python
• OpenCV
• NumPy
• Pandas
• Scikit-learn
• XGBoost
• scikit-image
• Pillow

---

Future Improvements

Possible improvements include:

• Faster FFT implementation
• Parallel feature extraction
• Additional frequency-domain descriptors
• Mobile deployment
• Hybrid classical + deep learning approach
• Larger and more diverse training dataset

---

License

This project was developed as part of an image classification assignment for educational purposes.