পাঠ ৩১ · ৩৫-এর মধ্যে · মডিউল ৪

Home / AI Courses / Computer Vision / Face recognition

Face recognition

Face recognition — detection, embedding, identification

৭ মিনিট পড়া মাঝারি · Intermediate Python কোডসহ

এই পাঠে যা শিখবেন

Face recognition pipeline 4 step
FaceNet triplet loss
ArcFace angular margin loss
Ethics ও privacy concern

১ · Face recognition vs detection

Face detection: "এই ছবিতে কোথায় মুখ?" — bounding box।
Face recognition: "এই মুখ কে?" — identity।
Verification: "এই দু'মুখ কি একই person?" — 1:1 match।
Identification: database-এ "এই মুখ কে?" — 1:N search।

কেন্দ্রীয় ধারণা

Recognition = "embedding learning"। Each face → 128-512D vector। Same person → close, different → far। Cosine similarity-এ comparison।

২ · Pipeline 4 step

Detection: face crop locate।
Alignment: 5-68 landmark detect → rotate/scale to canonical pose।
Embedding: CNN/ViT → vector।
Matching: compare embedding-এ cosine similarity → threshold।

৩ · Detection

Viola-Jones (২০০১): Haar cascade — classical, fast।
MTCNN (২০১৬): three-stage CNN। Face + landmark।
RetinaFace (২০১৯): single-stage, robust।
SCRFD (২০২১): efficient mobile।
YOLOv8-face: latest realtime।

৪ · Alignment

5-point: 2 eyes, nose, 2 mouth corner।
Affine transform → canonical (e.g., eyes horizontal, fixed scale)।
Eliminates pose variation।
Recognition accuracy boost ~5-10%।

৫ · FaceNet (২০১৫, Google)

Schroff et al. — embedding learning revolution।

Triplet loss:

$$\mathcal{L} = \max\big(0, \|a - p\|^2 - \|a - n\|^2 + \alpha\big)$$

$a$: anchor — base image।
$p$: positive — same person, different photo।
$n$: negative — different person।
$\alpha$: margin (typical 0.2)।
Push: anchor-positive close, anchor-negative far।

৬ · ArcFace (২০১৯)

Deng et al. — additive angular margin। Current SOTA।

Idea: classification head with angular margin between class।

$$\mathcal{L} = -\log \frac{e^{s \cos(\theta_{y_i} + m)}}{e^{s \cos(\theta_{y_i} + m)} + \sum_{j \ne y_i} e^{s \cos\theta_j}}$$

$\theta$: angle between feature and class center।
$m$: angular margin (e.g., 0.5)।
$s$: scale (e.g., 64)।
Inter-class margin enforce।

৭ · Recognition matching

Each face → embedding (e.g., 512-D)।
L2 normalize।
Compare: cosine similarity = $\cos\theta = a \cdot b$ (for unit vector)।
Threshold: > 0.6 same person, < 0.4 different (typical)।

Face recognition = "fingerprint of face"। Embedding vector = unique signature। Library compare — match found vs new identity।

Face recognition — detect, align, embed, match। Embedding-এ identity encoded।

৮ · Practical use

Python · DeepFace

# pip install deepface
from deepface import DeepFace

# Verification
result = DeepFace.verify(
    img1_path="person1.jpg",
    img2_path="person2.jpg",
    model_name="ArcFace",
)
print("Same person:", result["verified"])
print("Distance:", result["distance"])

# Identification
result = DeepFace.find(
    img_path="query.jpg",
    db_path="people_db/",
    model_name="ArcFace",
)
print("Match:", result)

# Face attribute (age, gender, race, emotion)
analysis = DeepFace.analyze(
    img_path="photo.jpg",
    actions=['age', 'gender', 'emotion'],
)
print(analysis)

DeepFace — high-level Python library। Multiple model support (FaceNet, ArcFace, VGG-Face, etc.)। Easy verification ও search।

৯ · Bias ও fairness

Buolamwini & Gebru (২০১৮): commercial face recognition dark-skin face accuracy 35% lower।
NIST FRVT: federal benchmark — bias documented।
Causes: training data demographic skew।
Mitigation: diverse dataset, fairness-aware loss, demographic-balanced eval।

১০ · Privacy ও regulation

EU GDPR: biometric data special protection।
EU AI Act: public space face recognition restricted।
US: San Francisco, Boston ban।
Bangladesh: Data Protection Bill emerging।
Best practice: consent, opt-in, secure storage।

১১ · Use cases — pros ও cons

Beneficial:

Phone unlock — Apple Face ID।
Photo organization — Google Photos।
Missing person search।
Building access (consensual)।
Bank KYC verification।

Concerning:

Mass surveillance।
Stalking enabling।
Wrongful arrest (false positive)।
Authoritarian misuse।

১২ · Bangladesh deployment

NID verification — bank, telecom।
Employee attendance (consent issue)।
Election verification (debate ongoing)।
Smartphone unlock।
Bangladesh-specific dataset rare — bias risk।

Face recognition extremely powerful এবং extremely dangerous। Engineer-এর responsibility — use case ethical evaluate। "Just because we can, doesn't mean we should."

ভাবনার প্রশ্ন

প্র ০১ Triplet loss-এ "hard negative mining" — কী ও কেন critical?

FaceNet training-এর key trick।

Naive triplet:

Random anchor + positive + negative।
Most triplet "easy" — anchor-negative already far।
Loss zero, no learning signal।

Hard negative:

Negative similar to anchor (high cosine similarity)।
Triplet loss > 0 — gradient flows।
Model learns to distinguish।

Mining strategies:

Hardest: closest negative। Effective but unstable।
Semi-hard: anchor-negative > anchor-positive but within margin। Stable।
Online mining: within batch, dynamic।

Empirical:

Naive: 60% LFW।
Hard mining: 99.6% LFW।
Massive difference।

Modern alternatives:

ArcFace — classification-based, no mining।
Sphere face — angular margin।
Contrastive — InfoNCE।

মূল উপলব্ধি: Sample selection ML training-এ underrated factor। Right hard examples >> more easy examples।

প্র ০২ ArcFace "angular margin" — geometry-এ কী ঘটছে? Cosine vs euclidean?

Embedding space geometry — face recognition-এর deep aspect।

Hyperspherical embedding:

L2 normalize → unit vector।
Embedding lives on hypersphere।
Distance = angular distance।
Cosine similarity = $\cos\theta$।

Why hyperspherical:

Magnitude irrelevant — only direction matter।
Image variation (lighting, etc.) magnitude-affecting, direction less।
Stable across condition।

Angular margin:

Each class — center on hypersphere।
Add margin $m$ to angle: $\cos(\theta + m)$।
Decision boundary "tight"।
Inter-class separation enforced।

Effect:

Same person — narrow cone around center।
Different person — angular gap > m।
Verification — single threshold across।

Vs Euclidean:

Euclidean — magnitude affects।
Cosine — only angle।
Image lighting/pose change — Euclidean fluctuate, cosine stable।

Modern face recognition:

SphereFace — multiplicative margin।
ArcFace — additive (cleaner)।
CosFace — variant।
MagFace — magnitude-aware।

মূল উপলব্ধি: Embedding space geometry matters। Right metric (cosine vs Euclidean) — significant accuracy gain।

প্র ০৩ Bangladesh-এ face recognition deploy — privacy law, consent, bias। Engineer-এর checklist?

Real Bangladesh ethical engineering question।

Legal landscape:

Bangladesh Data Protection Bill — pending।
ICT Act 2006 — broad provisions।
NID rule — biometric data government regulated।
EU GDPR equivalent emerging।

Consent framework:

Explicit, informed, specific।
Withdraw possible।
Children — parental consent।
Document consent log।

Bias mitigation:

Bangladesh-specific dataset — diverse skin tone।
Per-demographic accuracy test।
Public benchmark report।
External audit।

Security:

Embedding (not raw image) store — irreversible if salted।
Encryption at rest + transit।
Access log — audit trail।
Breach response plan।

Engineer checklist:

Use case ethical justify? (security vs convenience)।
Consent workflow design।
Bias evaluation per demographic group।
False positive/negative impact analysis।
Data retention policy।
Right to deletion implementation।
Third-party audit।
Public transparency report।

Use cases ranked:

✅ Phone unlock (personal, opt-in)।
✅ Photo album organize।
⚠ Office attendance (consent + alternative offered)।
⚠ Bank KYC (regulated, documented)।
❌ Public CCTV mass surveillance।
❌ Behavioral advertising।
❌ Without consent identification।

Bangladesh examples:

BD Election Commission — biometric verification।
bKash/Nagad — KYC।
Garment factory attendance — controversial।

Reading recommendation:

UNESCO AI ethics framework।
Microsoft AI principles।
Anthropic Constitutional AI।
Bangladesh CSA Cyber Security recommendation।

মূল উপলব্ধি: Engineer = ethical decision-maker। "Compliant" minimal bar; "ethical" higher। Bangladesh-এ AI ecosystem responsibility shape করার সুযোগ।

প্র ০৪ Face recognition adversarial attack — adversarial sticker, makeup। কী technique, defense কী?

Face recognition security — active research।

Attack types:

Evasion: mask/glasses fool detection।
Impersonation: appear as another person।
Adversarial patch: printed sticker disrupt।
Adversarial makeup: careful pattern।
Eye glasses: Sharif et al. 2016 — fool ResNet।

Adversarial example math:

Find perturbation $\delta$: $\| \delta \|_\infty < \epsilon$।
$\arg\max \mathcal{L}(f(x + \delta), y_t)$ where $y_t$ = target identity।
FGSM, PGD optimization।

Physical attack:

Cap, glass, sticker — robust to printing imperfection।
Multi-angle, multi-lighting test।
Demo: Wu et al., 2020 — IR-light attack।

Liveness detection:

Photo attack — passive countermeasure।
Active liveness — blink, smile।
3D depth (iPhone Face ID)।
Texture analysis।

Defense techniques:

Adversarial training: include adversarial examples।
Input preprocessing: JPEG compress, blur।
Detection: classify as adversarial।
Multi-modal: face + voice।
Continuous authentication: not single-shot।

Privacy-preserving counterattack:

Fawkes (UChicago, 2020) — protect own photo from training।
Cloak — adversarial perturbation।
Anti-FaceNet glasses।

Production reality:

Commercial system — multi-layer defense।
Apple Face ID — depth + IR + neural attestation।
Azure Face — liveness + behavior।
Banking — multi-factor essential।

মূল উপলব্ধি: Security ≠ accuracy। Adversarial robustness separate consideration। Bangladesh banking, government — defense critical।

অনুশীলন

DeepFace verify: Two photo same person কিনা — DeepFace।

Code section ৮-এ। ArcFace, FaceNet — multiple model try।

Embedding visualize: 5 person × 3 photo — embedding extract, t-SNE plot।

from sklearn.manifold import TSNE
embeddings = [DeepFace.represent(img)[0]['embedding'] for img in images]
tsne = TSNE(n_components=2).fit_transform(np.array(embeddings))
# Plot — same person should cluster

ভাবুন: Bangladesh smart attendance — face recognition vs biometric — কোনটা?

Fingerprint — physical contact, hygiene concern (post-COVID)। Face — contactless, fast। Privacy — both biometric। Local consent + opt-out essential।