Multi-State License Plate Recognition

PROJECT: Multi-State License Plate Recognition for US Tolling Systems

CHALLENGE: A highway tolling provider serving 5 US states had 95% accuracy in ideal conditions, but accuracy dropped to 62% in real-world adverse conditions (rain, fog, glare, dirt-covered plates, extreme angles). Their automated OCR couldn’t handle partial visibility or make contextual judgments on ambiguous characters. With 30% of images requiring manual review and high dispute rates from misread plates, they needed human annotators who could reason through edge cases.

OUR SOLUTION:
We deployed 70 specialized annotators trained on all 50 states’ license plate formats and edge case patterns. Our workflow combined contextual reasoning with multi-stage QA:
(1) First pass transcription with uncertainty levels,
(2) State format pattern validation,
(3) 30% peer review,
(4) QA lead verification on flagged cases.

Annotators didn’t just transcribe visible characters, they reasoned about likely characters based on state formatting rules, partial visibility, and common occlusion patterns. We delivered character-level bounding boxes in JSON format with confidence scores and uncertainty flags.

RESULTS: Improved accuracy from 62% to 84% in adverse conditions across 15,000+ edge case plates while maintaining 97.8% QA accuracy. Client’s toll collection dispute rate dropped 34% after model retraining. Automated system now flags only 8% of images (down from 30%), reducing operational costs. Performance improvement enabled deployment in 2 additional states previously considered too challenging.

DELIVERABLES:
800,000+ annotated license plates
JSON format with character bounding boxes
97.8% QA accuracy