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.