Building a Fixed-Length CAPTCHA OCR Model With Multi-Head Classification

This article documents the design of a lightweight OCR system built to solve fixed-length numeric CAPTCHAs for authorized internal automation workflows. Instead of using a standard CRNN + CTC architecture, the author built a shared CNN backbone with six independent classification heads and learnable position embeddings, achieving 100% held-out accuracy with roughly 4,000 training samples while improving training stability, inference speed, and debuggability