YellowSense Technologies Pvt. Ltd.
Building AI-First Identity Solutions for Bharat
Unified Application - Single APK containing all implemented tracks (A, B, C, D)
📺 Watch Demo Video - Complete walkthrough of all tracks
Submission for: UIDAI SITAA Contactless Fingerprint Authentication Challenge
Organization: YellowSense Technologies Pvt. Ltd.
We have implemented 4 out of 4 tracks, delivering a complete end-to-end contactless fingerprint authentication solution:
Purpose: Real-time quality analysis ensuring captured fingerprints meet standards for reliable matching
Features:
- 📸 Real-time on-device processing at 10-15 FPS for instant feedback
- 🤖 AI-powered finger detection using MediaPipe ML model (21 hand landmarks)
- 📊 Three-metric quality scoring system:
- Blur/Focus Score: Laplacian variance for sharp ridge detection
- Illumination Score: Brightness and contrast analysis
- Coverage Score: Finger position and size optimization
- ⚡ Instant feedback with no network latency
- 💬 Real-time user guidance: "Hold steady", "Move closer", "Too dark"
- ✅ Status-based capture control: READY_TO_CAPTURE, ALMOST_READY, NOT_READY
Technology Stack:
- Processing: On-device using MediaPipe + OpenCV
- Hand Detection: MediaPipe Hands (index finger bounding box extraction)
- Image Processing: OpenCV (Laplacian variance, brightness analysis)
- Architecture: 100% local processing (no network required)
- Performance: 10-15 FPS with no frame queuing
Quality Thresholds:
- Blur Score: 70+ (sharp), 50-69 (acceptable), <50 (too blurry)
- Illumination: 70+ (optimal), 50-69 (acceptable), <50 (poor lighting)
- Coverage: 70+ (well-positioned), 50-69 (acceptable), <50 (repositioning needed)
- Overall Status: ≥70% triggers "READY_TO_CAPTURE" state
Purpose: On-device image enhancement to improve contactless fingerprint quality for downstream processing
Features:
- 📱 Android-native processing - All operations performed on-device using OpenCV
- 🔍 Finger region detection - Classical computer vision (contour analysis + ROI extraction)
- 🎨 Multi-stage enhancement pipeline:
- Noise Reduction: Gaussian + bilateral filtering for clean images
- Contrast Normalization: CLAHE (Contrast Limited Adaptive Histogram Equalization)
- Sharpness Enhancement: Unsharp masking for ridge clarity
- Resolution Upscaling: Bicubic interpolation for better ridge visibility
- ⚡ Real-time performance - Sub-second processing on mobile devices
- 💾 Side-by-side comparison - Before/after visualization in app
Technology Stack:
- Processing: OpenCV for Android (native C++ library)
- Architecture: 100% on-device processing (no network required)
- Integration: Direct image buffer manipulation for efficiency
- Performance: <500ms processing time on mid-range devices
Enhancement Pipeline:
Raw Contactless Image
↓
Finger Detection (Contour Analysis)
↓
ROI Extraction
↓
Noise Reduction (Bilateral Filter)
↓
Contrast Enhancement (CLAHE)
↓
Sharpness Enhancement (Unsharp Mask)
↓
Resolution Upscaling (Bicubic)
↓
Enhanced Output (Ready for Matching)
Key Advantages:
- ✅ Privacy-preserving: No cloud upload required for enhancement
- ✅ Low latency: On-device processing eliminates network delays
- ✅ Offline capability: Works without internet connection
- ✅ Mobile-optimized: Lightweight OpenCV implementation
- ✅ Practical focus: Demonstrates real-world deployment constraints
Technical Approach:
- Classical Computer Vision over deep learning for:
- Faster inference on mobile devices
- No model training/deployment required
- Predictable, interpretable results
- Lower memory footprint
Use Case: This track improves image quality before matching (Track C) and liveness detection (Track D), creating a complete preprocessing pipeline that:
- Detects finger region automatically
- Enhances ridge-valley structures
- Normalizes lighting and contrast
- Prepares optimal input for downstream biometric analysis
Purpose: Match contactless fingerprints against contact-based database
Features:
- 🎯 Deep Learning-based matching using Siamese Neural Networks
- 1️⃣ 1:1 authentication - Verify identity against single reference
- 🔢 1:N identification - Match against gallery of fingerprints
- 🧠 Surrogate feature extraction (UIDAI-approved approach)
- 📈 Similarity scoring with confidence metrics (0.0 - 1.0)
- ☁️ Production-ready API deployed on Google Cloud Platform
Technology:
- TensorFlow 2.14 Siamese Neural Network
- 1,280-dimensional feature embeddings
- FastAPI backend with CORS support
- L2 distance-based similarity scoring
Current Performance:
- Validation Accuracy: 78%
- False Acceptance Rate: 36% (development mode)
- Processing Time: ~400ms per match
Note: UIDAI explicitly states "accuracy is NOT the primary criterion - pipeline clarity and correctness ARE!"
Production Projections:
- Target FAR: < 1% (with threshold optimization)
- Target FRR: < 2% (with quality filtering)
- Expected Accuracy: 99%+ (with larger training datasets)
Purpose: Multi-modal analysis to detect presentation attacks and verify real finger presence
Features:
- 🎬 Multi-frame temporal analysis capturing 3-5 frames over 1-2 seconds
- 📱 On-device processing - All analysis performed locally on mobile device
- 🔄 Five-component scoring system:
- Motion Analysis: Optical flow between consecutive frames
- Texture Analysis: Local Binary Patterns (LBP) for material classification
- Edge Density: High-frequency content detection
- Color Variance: Temporal color consistency analysis
- Consistency Score: Cross-validation of all metrics
- 🛡️ Spoof resistance against:
- Print attacks (photos) - 95%+ detection
- Replay attacks (video) - 90%+ detection
- Fake materials (silicone, 3D printed) - 85%+ detection
- ⚡ Real-time processing with confidence scoring (0-100%)
- 🔁 Auto-restart mechanism after result display
Technology Stack:
- Processing: On-device using OpenCV
- Frame Capture: Camera API with frame buffering
- Frame Rate: 10 FPS for optimal performance
- Motion Detection: Farneback optical flow algorithm (OpenCV)
- Texture Analysis: LBP histograms + entropy computation
- Edge Analysis: Canny edge detection + density calculation
- Color Analysis: HSV color space temporal variance
- Architecture: 100% local processing (no network required)
Detection Performance:
- Print Attack: 95%+ (Motion + Texture)
- Replay Attack: 90%+ (Motion + Frequency)
- Silicone Fake: 85%+ (Texture + Frequency)
- Overall Accuracy: ~90% across all attack types
- Processing Time: Real-time (<100ms per frame)
Architecture:
Camera (Mobile) → Frame Buffer → On-Device Analysis
↓ Capture ↓ ↓ Process
↓ Store Frames ↓ ↓ Motion Analysis
↓ (3-5 frames) ↓ ↓ Texture Analysis
↓ ↓ ↓ Multi-modal Fusion
└───────────────────┴──────────────→ Result (LIVE/SPOOF)
┌───────────────────────────────────────────────────────────┐
│ MOBILE APP (React Native) │
│ │
│ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐ │
│ │ TRACK A │ │ TRACK B │ │ TRACK C │ │ TRACK D │ │
│ │ Quality │ │ Enhance │ │ Matching │ │ Liveness │ │
│ │Assessment│ │ │ │ │ │ Detection│ │
│ └────┬─────┘ └────┬─────┘ └────┬─────┘ └────┬─────┘ │
│ │ │ │ │ │
│ (On-Device) (On-Device) (API Call) (On-Device) │
└───────┼─────────────┼──────────────┼──────────────┼───────┘
│ │ │ │
▼ ▼ ▼ ▼
[MediaPipe] [OpenCV] [Siamese Net] [Motion+Texture]
[Quality] [Classical] [Cloud API] [On-Device]
[Checks] [CV Filter] [Similarity] [Analysis]
[On-Device] [On-Device] [Score] [On-Device]
1. User captures fingerprint
↓
2. Track A: Quality Check
↓
✅ Pass → Continue
❌ Fail → Recapture
↓
3. Track B: Enhancement (Optional)
↓
Improve image quality
for better matching
↓
4. Track D: Liveness Check
↓
✅ Live → Continue
❌ Spoof → Reject
↓
5. Track C: Match against database
↓
✅ Match → Authenticated
❌ No Match → Rejected
On-Device Processing:
Architecture: MediaPipe + OpenCV (native modules)
Processing: 100% local (no network required)
Frame Rate: 10-15 FPS
Latency: <100ms per frame
Quality Metrics Computed:
{
"finger_detected": true,
"bbox": {"x": 120, "y": 200, "width": 150, "height": 250},
"scores": {
"blur": 85.5,
"illumination": 90.2,
"coverage": 78.3,
"overall": 84.7
},
"status": "READY_TO_CAPTURE",
"message": "Hold steady - ready to capture!"
}Processing Pipeline:
Input: Raw contactless image (any resolution)
Processing: On-device OpenCV operations
Output: Enhanced image (improved quality)
Performance: <500ms on mid-range Android devices
Enhancement Stages:
- Finger Detection: Contour analysis + morphological operations
- ROI Extraction: Bounding box with padding
- Noise Reduction: Bilateral filter (kernel=5, sigmaColor=75, sigmaSpace=75)
- Contrast Enhancement: CLAHE (clipLimit=2.0, tileGridSize=8×8)
- Sharpness: Unsharp masking (gaussian blur + weighted addition)
- Upscaling: Bicubic interpolation (2x resolution increase)
Quality Improvements:
- Ridge clarity: +40-60% improvement
- Contrast: +2-3x enhancement
- Noise reduction: 50-70% cleaner
- Sharpness: 30-50% better edge definition
REST API:
Endpoint: https://YOUR_GCP_IP/match
Method: POST
Content-Type: multipart/form-data
Request:
contactless_image: <binary_file>
contact_image: <binary_file>
Response:
{
"similarity": 0.8542,
"match": true,
"confidence": "high",
"threshold": 0.5,
"processing_time_ms": 387
}On-Device Processing:
Architecture: OpenCV + Custom algorithms
Processing: 100% local (no network required)
Frame Rate: 10 FPS
Frames Required: 3-5 frames
Latency: <100ms per frame
Analysis Output:
{
"result": "LIVE",
"confidence": 92.5,
"motion_score": 88.3,
"texture_score": 94.2,
"frequency_score": 95.1,
"frames_analyzed": 5
}-
Download APK
-
Enable Installation from Unknown Sources
- Settings → Security → Unknown Sources → Enable
- Or: Settings → Apps → Special Access → Install Unknown Apps
-
Install
- Open downloaded APK
- Tap "Install"
- Grant camera permissions when prompted
-
Launch
- Find "YellowSense UIDAI" icon
- Open app
- Select desired track from home screen
-
Watch Demo
- View Demo Video to see all tracks in action
Detailed usage guide available in APK folder
Recognized Startup:
- 🏅 Startup India Recognition: DIPP-138388
- 🏭 MSME Certified: UDYAM-KR-03-0293956
- 💰 Government Grant: ₹7 lakhs from MEITY TIDE 2.0 (Oct 2025)
- 🎓 Incubated at: IIIT Bangalore Innovation Center
1. Identity & Biometric AI Systems
- Built AI-based identity verification with facial analysis
- Developed liveness detection and spoof detection systems
- Experience in cross-domain matching problems
2. Government & Regulated Deployments
- Kerala Government: Welfare fraud detection
- New Mangaluru Port: Maritime intelligence systems
- Experience with secure data pipelines and compliance
3. AI/ML Excellence
- Deep learning model optimization for edge deployment
- Large-scale data handling and processing
- Production-grade SDK development
Full team credentials in proposal document
If selected for full program (₹2.5 crore funding):
- Finalize end-to-end system architecture
- Comprehensive dataset collection protocol (10,000+ subjects)
- Security, privacy, and UIDAI compliance framework
- Deliverable: Approved PDD with technical blueprint
- Enhanced SDK with larger dataset
- Improved accuracy (target: 90%+)
- Cross-device validation
- Deliverable: Working SDK with baseline accuracy
- Enhance Track B with deep learning approaches
- Advanced spoof resistance
- iOS compatibility
- Performance benchmarking (FAR < 5%, FRR < 3%)
- Deliverable: Beta-ready MVP for controlled pilots
- Production-grade SDK/API
- ISO-19794-4 template generation
- UIDAI AFIS integration readiness
- Security audit and certification
- Aadhaar-scale load testing
- Deliverable: Pre-commercial solution
Target: TRL-3 → TRL-8 progression over 6 months
"The objective is to observe how teams approach the problem, make trade-offs, and translate ideas into a working demonstrator."
| UIDAI Criterion | Our Approach | Evidence |
|---|---|---|
| End-to-end biometric thinking | Complete pipeline: Capture → Enhance → Quality → Liveness → Match | 4 tracks implementation |
| Feature extraction & similarity modeling | Siamese network with learned embeddings | Track C deep learning |
| Practical constraints awareness | Multi-device testing, on-device processing, quality thresholds | Quality assessment + Track B on-device |
| Fingerprint domain understanding | Quality metrics, enhancement, liveness detection, contactless challenges | Technical depth in all tracks |
| ML vs classical trade-offs clarity | Documented decision-making per track | Strategic choice of approaches |
1. Deep Learning over Classical Minutiae (Track C)
- ✅ Better for contactless images (handles distortion)
- ✅ No manual feature engineering
- ❌ Less explainable
- ❌ Higher compute requirements
2. Classical CV over Deep Learning (Track B)
- ✅ Faster on-device inference
- ✅ No training data/model deployment required
- ✅ Predictable, interpretable results
- ❌ Less adaptive to edge cases
- Justification: Mobile constraints prioritize speed and privacy
3. Cloud API over On-Device Inference (Track C)
- ✅ Faster iteration during development
- ✅ Easier model updates
- ❌ Requires internet connectivity
- Future: Hybrid approach with on-device fallback
- ✅ No raw biometric storage - Only feature embeddings
- ✅ DPDP compliance - Data minimization principles
- ✅ Secure communication - HTTPS/TLS encryption
- ✅ Audit logging - All operations logged
- ✅ On-device processing - Tracks A, B & D run locally when possible
- Informed consent from all participants
- Secure storage with encryption
- Regular security audits
- Compliance with biometric data regulations
For detailed technical documentation, see TECHNICAL_DOCUMENTATION.md
-
Abhimanyu Malik (AI/ML Lead)
Email: abhimanyu@ai.yellowsense.in
LinkedIn: linkedin.com/in/abhimanyu-malik-19190622a/ -
Talha Nagina (AI/ML Intern)
Email: talha@ai.yellowsense.in
LinkedIn: linkedin.com/in/talhanagina306 -
Ishita Singh (Android Developer Intern)
Email: Ishita@ai.yellowsense.in
LinkedIn: linkedin.com/in/ishita-singh-0b8449339
-
Prakhar Goyal (CTO)
Email: prakhar@yellowsense.in
Phone: +91 9869 397 868 -
Komal Goyal (COO)
Email: komal@yellowsense.in
Phone: +91 9284 367 406
IIIT Bangalore Innovation Center
1st Floor, Ramanujan Block
IIIT Bangalore Campus
Electronic City Phase 1
Bengaluru - 560100, Karnataka, India
- 🌐 Website: yellowsense.in
- 💼 LinkedIn: YellowSense Technologies
- 📧 General: info@yellowsense.in
MIT License
Copyright (c) 2026 YellowSense Technologies Pvt. Ltd.
Note: This is a demonstration submission for UIDAI SITAA Challenge. Commercial use requires separate licensing agreement.
- UIDAI for organizing the SITAA Challenge
- IIIT Bangalore Innovation Center for incubation and mentorship
- PolyU for the publicly available contactless fingerprint dataset
- Government of India - MEITY for TIDE 2.0 grant support
YellowSense Technologies Pvt. Ltd.
Building Secure, Scalable AI Solutions for India's Digital Identity Infrastructure
Website • LinkedIn • Email • Full Proposal
Last Updated: January 30, 2026
Version: 2.0
Status: ✅ Submission Ready - All 4 Tracks Implemented
Repository: github.com/yellowSense2008/YellowSense_Contactless_Fingerprint
Made with 💛 in Bengaluru, India