IndustryBRAC Bank PLCWork period · 2025-2026Enterprise Multi-Camera Network

BRAC CIN: Face Recognition & Identity Review

Identity-aware banking surveillance using InsightFace embeddings, FAISS similarity search, whitelist and blacklist categories, unknown-person enrollment, and a review workflow for registry quality.

Module of BRAC CIN; presented as a separate case study.

BRAC CIN: Face Recognition & Identity Review project workflow and system architecture
Project workflow and system architecture overview. Open full-size diagram

Workflow at a glance

  1. Detect faces

    Prepare visible face crops from camera frames.

  2. Compare embeddings

    Search registered embeddings for similarity.

  3. Assign review state

    Separate registry matches and unknown faces.

  4. Review registry

    Store evidence for identity review workflows.

Project facts & reported results

InsightFace / ArcFace
Embedding

Implementation fact

FAISS cosine similarity
Search

Implementation fact

Whitelist, blacklist, unknown review
Workflow

Implementation fact

Project Overview

Problem Statement

Long-running face registries can accumulate low-quality or incorrect enrollments, especially under low light, occlusion, and changing ATM camera views.

Approach & Methodology

Engineered multi-camera pipeline using InsightFace for face detection and ArcFace for 512-dimensional embedding generation. Built an interactive web review panel for security officers to validate WL/BL matches and audit face registry quality. Implemented automated registry pollution detection scripts and a background daemon for reliable alert posting.

Outcome & Scope

Combines real-time matching with a human review queue, registry audit tools, and decoupled alert delivery for maintainable identity operations.

Technical Challenges & Solutions

1

Face Registry Pollution

Poor quality face images from blurry or low-light video clips polluted the face database, causing incorrect matches over time.

Solution

Developed automated registry pollution audit scripts and created a web-based review dashboard for security personnel to audit and re-enroll clean face samples.

System Architecture

Two-process surveillance system with real-time face embedding and web audit dashboard

Key Components:

Face Detection & Embedding Engine

InsightFace + ArcFace pipeline extracting 512-dimensional vectors from camera frames

WL/BL Match Evaluator

Compares face vectors against whitelist/blacklist registry using cosine distance

Web Review Panel

Interactive HTML/Flask dashboard for security officers to review matches and submit feedback

Registry Audit & Recovery Service

Tooling for identifying corrupted embeddings and restoring clean registry state

Alert Sender Daemon

Asynchronous queue manager posting verified face events to CIN API

Key Features & Capabilities

Real-time face detection & tracking across multiple ATM cameras

InsightFace detection + ArcFace 512-dimensional embedding extraction

Whitelist/Blacklist (WL/BL) match classification with customizable thresholds

Web-based interactive review dashboard for security audit and match feedback

Registry pollution audit & automated face registry reset/recovery scripts

Asynchronous alert daemon dispatching confirmed matches to CIN API

Technologies & Tools

InsightFace
ArcFace 512d
Python
SQLite
OpenCV
Flask
CIN API
Multi-Camera