Face Recognition for Smart Voting
A computer vision pipeline built to make voting tamper-proof. One face, one vote, verified biometrically.
An IEEE-presented study applying computer vision to India's duplicate-vote problem. A Haar Cascade and LBPH hybrid pipeline that biometrically authenticates voters before ballot access. Presented at Smart GENCON 2021, IEEE Pune Section.
At a glance
The Story
The Problem — India's Duplicate Vote Challenge
In Indian elections, manual voter verification struggles to prevent impersonation and duplicate balloting at scale. The question was whether face recognition could replace that manual gate. A biometric check before the ballot is cast would enforce one-person-one-vote without human bottlenecks.
Surveying the Landscape — 14 Face Recognition Methods
A systematic review of 14 published studies benchmarked approaches across Haar Classifiers, LBPH, EigenFace, FisherFace, PCA, LDA, and deep learning (CNN, YOLOv3). The goal was to identify which hybrid would be most reliable for single-voter, real-time deployment — not just peak theoretical accuracy.
| Method | Application | Notable Result |
|---|---|---|
| Haar + LBPH + FisherFace | Visual-assist wearable | LBPH 93.33% · Fisher 86.67% |
| PCA + LDA + Neural Network | Face recognition | 99.5% on YALE dataset |
| FisherFace | Face recognition | 100% train · 93% test |
| PCA Euclidean Distance | Biometric FR | 98–100% on ORL dataset |
| Haar + EigenFace (PCA) | Face recognition | 91.67% multi-face similarity |
| YOLOv3 + Haar Cascade | Face mask detection | 90.1% |
| Haar + Keras Adam | Gender identification | Train 96% · Val 90% |
| CNN + SoftMax classifier | Face recognition | ~100% train · 90% test |
| Haar + CNN (MobileNetsV2) | Drowsiness detection | ~87% |
| Haar + LBPH | Smart attendance | 81.6% |
| Haar + TF-IDF | Face detection in dim light | Works well in grey-scale |
| Haar + Instagram selfie (Haar) | Social media analysis | 71.48% (high false-pos) |
Test accuracy across surveyed face recognition methods
Building the Pipeline — Haar Detection + LBPH Recognition
The system pipeline flows through four stages: (1) Pre-processing — grayscale conversion and noise reduction to boost algorithm accuracy; (2) Face Detection — Haar Cascade classifier using edge, line, and centre-surround Haar-like features, tuned via scaleFactor (image scale step) and minNeighbors (false-positive suppression); (3) Feature Extraction — Haar-like features drawn from a 24×24 sub-window (~162,000 possible features); (4) Face Recognition — LBPH histogramming the local texture to match against trained voter identities. All implemented in Python with OpenCV.
| Pipeline Stage | Technique | Purpose |
|---|---|---|
| Pre-processing | Grayscale conversion | Haar-like features require greyscale; reduces noise |
| Face Detection | Haar Cascade Classifier | Detects face region using scaleFactor & minNeighbors params |
| Feature Extraction | Haar-like features (24×24 window) | ~162,000 possible edge, line, centre-surround features |
| Face Recognition | LBPH Face Recognizer | Histogram-based texture matching against voter database |
| Alternative (evaluated) | EigenFace — PCA-based | Reduces face-space dimensions; best for controlled lighting |
| Alternative (evaluated) | FisherFace — PCA + LDA | Separates individuals across classes; handles illumination variation |
Results — What Single vs Multi-Face Testing Revealed
The hybrid Haar+LBPH model reliably distinguished facial from non-facial images and correctly authenticated individual voters. The critical finding was an edge case: multi-face images degraded recognition confidence, while single-face images produced stable results. For a Smart Voting System — where voters stand one at a time — this boundary condition validates the deployment model. LBPH's 93.33% result in the closest analogous domain (wearable visual-assist device) confirmed it outperformed FisherFace (86.67%) for real-world texture variation.
| Scenario | Outcome | Implication |
|---|---|---|
| Single-face image input | Reliable recognition ✓ | Validates Smart Voting deployment model |
| Multi-face image input | Reduced confidence / unordinary results | Boundary condition: one voter at a time enforced |
| Non-facial image input | Correctly rejected ✓ | System doesn't false-positive on background objects |
| LBPH vs FisherFace | 93.33% vs 86.67% (wearable-device domain) | LBPH selected as primary recognizer |
Presented at IEEE Smart GENCON 2021
The paper was presented at the 2021 International Conference on Smart Generation Computing, Communication and Networking (SMART GENCON), 29–30 October 2021, organised by IEEE Pune Section. The Haar Cascade + LBPH Smart Voting System was accepted as a contribution to the emerging field of biometric security infrastructure for democratic systems.
IEEE Smart GENCON 2021 — Presentation Certificate
Published Papers
Optimal Face Recognition System using Haar Classifier
IEEE Smart GENCON 2021 — International Conference on Smart Generation Computing, Communication and Networking · 2021
Soham Wattamwar · Rishikesh Mate · Parth Rainchwar · Shamla Mantri · Geeta Sorate
A hybrid face recognition pipeline using Haar Cascade for face detection and LBPH for identity recognition, designed for a Smart Voting System that biometrically authenticates voters before ballot access. The study benchmarks Haar-based approaches against EigenFace, FisherFace, CNN, and PCA-based methods across 14 prior works before implementing and evaluating the Haar+LBPH combination.
93.33%
LBPH Accuracy
86.67%
FisherFace Accuracy
14
Papers Reviewed
IEEE Smart GENCON
Conference