The aim of this project is to identify the source scanner device used to scan a document or image by analyzing the unique patterns or artifacts left behind during the scanning process.
Each scanner (brand/model) introduces specific noise, texture, or compression traces that can be learned by a machine learning model.
This project is important in forensic investigations, copyright authentication, and document verification tasks.
- Description: Determine which scanner was used to forge or duplicate legal documents.
- Example: Detect whether a fake certificate was created using a specific scanner model.
- Description: Identify the source of printed and scanned images to detect tampering or fraudulent claims.
- Example: Differentiate between scans from authorized and unauthorized departments.
- Description: Ensure scanned copies submitted in court/legal matters came from known and approved devices.
- Example: Verify that scanned agreements originated from the company’s official scanner.
By the end of this project, students will:
- Understand the concept of source device identification.
- Extract scanner-specific features such as noise patterns, frequency domain signals, and artifacts.
- Train a classification model to distinguish among multiple scanners.
- Evaluate model accuracy and visualize feature importance.
- Optionally deploy a simple app to upload and identify the scanner source.
- Source: Kaggle
(Additional manually collected samples may be added if needed.)
(Insert system architecture diagram here)
- Manually scan sample images using multiple scanners.
- Assign proper labels based on source device.
- Resize, denoise, and convert to grayscale if needed.
- Normalize pixels and remove non-artifact content.
- Extract noise patterns using filters (e.g., Wavelet, FFT).
- Compute PRNU, texture descriptors, and edge patterns.
- Train classifiers such as:
- CNN (if using deep features)
- Random Forest, SVM (if using extracted features)
- Evaluate performance on validation set.
- Upload an image → Return probable scanner model.
- Optional: Display confidence score and feature map.
- Collect scanned document samples from different scanner devices (minimum 3–5 models/brands).
- Create a labeled dataset (e.g.,
scanner_model,file_name, etc.). - Analyze basic image properties such as resolution, format, and color channel.
- Perform image preprocessing:
- Resize all images to a fixed dimension.
- Convert to grayscale if needed.
- Denoise (optional).
- Normalize and structure the dataset for model training.
- Extract hand-crafted features such as:
- Noise patterns.
- Frequency domain features (e.g., FFT).
- Texture descriptors (e.g., LBP).
- Visualize differences between scanner outputs (e.g., noise maps).
- Train baseline ML models (e.g., Logistic Regression, SVM, Random Forest).
- Evaluate using accuracy and confusion matrix.
- Log performance and identify limitations of hand-crafted features.
- Build and train a CNN model on the raw image dataset.
- Use image augmentation for generalization (brightness, rotation).
- Tune hyperparameters and track training curves (accuracy/loss).
- Evaluate model performance using accuracy, F1-score, and confusion matrix.
- Apply explainability tools such as SHAP or Grad-CAM to visualize how the model identifies scanner-specific patterns.
- Create a simple UI using Streamlit (or any frontend):
- Upload scanned image.
- Get predicted scanner brand/model with confidence score.
- Log predictions and allow download of results.
- Final documentation and formatting.
- Add system architecture, training results, model comparison, and screenshots.
- Prepare final presentation slides and demonstrate the working model.
- Data collected and labeled correctly.
- Feature engineering and modeling completed.
- UI integration and testing done.
- Classification accuracy above baseline (ideally >85%).
- Model should distinguish between at least 3–5 scanners.
- Robustness to image format and scan resolution.
- Clear explanation of methodology.
- Model performance charts (accuracy, confusion matrix).
- Insights into feature importance (explainability).
- Accuracy: Correct scanner prediction rate.
- Precision: Correct predictions for each scanner class.
- Recall: Ability to detect all true scanner outputs.
- F1-Score: Balance of precision and recall.
- Confusion Matrix: Visualization of misclassification between scanner types.
- Use SHAP or Grad-CAM (if CNN used) to show key areas.
- Validate that the model is not biased toward scan brightness or layout.
This project aims to bridge forensic analysis and machine learning by identifying scanner device signatures embedded in scanned images.
Through advanced feature extraction, classification, and explainable AI, the TraceFinder system can significantly aid digital forensics, legal investigations, and document authenticity verification.