High-Accuracy GCC ANPR Engine

I need a license-plate recognition engine that can reliably read Qatar, UAE, and Saudi Arabia plates when they appear in either still images or live video streams. High accuracy is non-negotiable—I am targeting production-level performance suitable for enforcement and analytics, so false positives must be minimal and confidence scores exposed.

To succeed you will design, train, and package a full ANPR pipeline that:
• detects the plate region in diverse lighting and angles,
• recognises Arabic and Latin characters as well as the specific colour codes and icons used across the three countries, and
• delivers results fast enough for real-time video (traffic-camera frame rates) while also accepting batch image inputs.

Please outline your preferred tech stack (OpenCV, YOLOv8, TensorFlow, custom CNNs, etc.), any data you already hold for these plate formats, and what extra data or annotations you would need from me. Your deliverable should include the trained model, inference code (Python preferred), and a simple REST or gRPC interface so we can drop it straight into our existing pipeline.

Acceptance will be based on:
• ≥ 95 % character recognition accuracy on a blind test set I will provide,
• ≤ 100 ms average inference per HD video frame on an NVIDIA T4 or equivalent, and
• clear installation and usage documentation.

If you have previously deployed ANPR systems for GCC plates—or can demonstrate equivalent multilingual OCR work—please share examples. I’m ready to move quickly once I see a solid approach.