(SYS.05) — PERSONAL PRODUCT — BROADCAST AUTOMATION & COMPUTER VISION
Jeizi OCR Controller
Real-time esports OCR and broadcast-telemetry engine for live tournament overlays.
(01) — System Overview
PROJECT
CONTEXT.
A high-performance .NET 8 Windows desktop application that converts live video broadcast streams into structured telemetry in real time. Features multi-source capture (OBS Virtual Camera, capture cards, desktop), OpenCV preprocessing filters, local Tesseract OCR, noise rejection stabilization, and atomic TXT/JSON/HTTP output.
The Operational Bottleneck / Problem
Esports broadcasts require live scoreboard data to drive stream overlays, but manual operator data-entry is error-prone and official game publisher APIs are frequently unavailable or restricted.
Engineering Solution
Engineered an automated desktop computer vision system that captures live video frames, extracts critical visual regions, stabilizes OCR results, and dispatches clean data to overlays in real time.
(02) — Architecture & Pipeline
SYSTEM
ARCHITECTURE.
DirectShow and desktop capture pipelines feeding shared frame buffers to OpenCvSharp image filters and Tesseract 5.2 OCR, stabilized via consecutive-match queues and written atomically to disk or network endpoints.
Data & Execution Pipeline
(03) — Technical Focus
ENGINEERING
HIGHLIGHTS.
Hierarchical Scene & Virtual Source View Architecture
Designed a Profile → Scene → Source View → OCR Field hierarchy allowing multiple isolated regions to process off a single capture stream, eliminating duplicate frame grabs.
Computer Vision Preprocessing Pipeline
Integrated OpenCV grayscale, contrast adjustment, Otsu thresholding, fixed thresholding, chromakey extraction, and morphological operations to clean broadcast graphics before OCR.
Noise Rejection & Consecutive-Match Stabilization
Engineered recognition stabilization requiring consecutive matching frames and whitespace sanitization, preventing bad video frames from corrupting live broadcast overlays.
Atomic Multi-Format Broadcast Dispatch
Implemented atomic batched TXT, JSON (`live_ocr.json`), and HTTP controller push pipelines to prevent partial reads or file-lock collisions with OBS and browser overlays.
(04) — Implemented Capabilities
CORE
FEATURES.
Multi-source capture: displays, windows, webcams, capture cards, OBS Virtual Camera
Scene-based execution: only the active tournament scene consumes OCR processing
Interactive region editor with draggable handles directly over live previews
Specialized recognition modes for numbers, timers, decimals, and general text
Esports tournament templates (kills, towers, gold, game timers, player statistics)
OpenCV image filters: Otsu thresholding, contrast, chromakey, dilation/erosion
Atomic TXT file writing for direct OBS text source integration
Consolidated `live_ocr.json` output for browser-based broadcast graphics
Direct HTTP telemetry streaming to external tournament control servers
Automated 43-suite regression test framework (`dotnet run -- --test`)
(05) — Technology Stack
STACK
ARCHITECTURE.
Runtime & Language
Computer Vision & OCR
Telemetry & Integration
Packaging & QA
(06) — Problem Solving
TECHNICAL
CHALLENGES.
Preventing Stream Jitter from Transient Video Artifacts
Built a consecutive-frame confirmation algorithm that rejects one-off frame anomalies, maintaining last-known-good values until new readings achieve statistical confidence.
High Performance with Multiple Scoreboard Fields
Reused a single shared frame capture buffer across all active source views and restricted OCR processing strictly to the active production scene.
(07) — Reliability & Governance
SECURITY & TESTING.
Privacy & Security Model
Entirely local desktop execution with zero cloud telemetry requirements.
Quality Assurance & Tests
43 automated component test suites verifying coordinate mapping, sanitization, scheduler timing, and atomic IO.
EXPLORE MORE WORK