Biometrics Identification System

Project Context

This project presents a murder investigation scenario where the only evidence is a fingerprint, and the goal is to identify the perpetrator from a database of 100 fingerprints. The system implements a comprehensive multimodal biometric recognition pipeline combining fingerprint and iris analysis

Fingerprint Recognition

For fingerprint recognition, a custom FP enhancement pipeline was developedusing gradient-based segmentation, orientation estimation, and Gabor filtering. The minutiae detection algorithm identifies ridge terminations and bifurcations using 8-neighborhood analysis, followed by keypoint matching and global alignment techniques.

Fingerprint preprocessing pipeline

Minutiae matching visualization

Using a custom composite similarity metric, a similarity table was created and the below 8 fingerprints were returned. Clearly, these fingerprints are all the same, therefore our database must have been tampered with. This is where the Iris Recognition pipeline comes in.

Top 8 fingerprint matches returned by the system

Iris Recognition

Similar to before, a custom preprocessing & enhancement system was applied to the raw iris images. Iris images underwent segmentation by means of Hough transform, and Daugman's rubber sheet model was used to normalise/unwrap the irises.

Iris preprocessing pipeline

A triplet loss encoder was used to encode the irises into robust feature embeddings that can effectively distinguish between different individuals' irises. Finally, a cosine similarity table was created and the top-k matches were returned.

Triplet loss encoder architecture

Multimodal Fusion

A weighted score-level fusion was used to bring together both the Fingerprint and Iris identification systems and provide a confident conclusion. This complete system allows for a more robust and accurate biometric identification system, while also handling edge-cases where one modality might fail.

Multimodal fusion results

Evaluation & Results

Comprehensive evaluation includes FAR/FRR curves, ROC analysis, DET curves, and CMC curves for both verification and identification scenarios. The system demonstrates improved accuracy through multimodal fusion compared to individual biometric systems, ultimately solving the forensic investigation by identifying the perpetrator through combined fingerprint and iris evidence.