Optical Character Recognition using VL

This project is an AMD ROCm based optical character recognition pipeline for the LabLab AMD AI Academy Mini Challenge 2. It uses Qwen2.5-VL-3B-Instruct with PyTorch and Hugging Face Transformers to read text from license plates and traffic signs. The pipeline handles US plates with state names and slogans, Chinese plates, stop signs, speed limit signs, advisory speed plaques, and multi-line warning or work-zone signs. Its instructions and post-processing are designed to keep only registration characters on US plates, preserve relevant characters on Chinese plates, and transcribe visible sign text in reading order without adding explanations or inferred characters. The solution accepts PNG, JPEG, and TIFF images through the required command-line interface and writes one JSON response per image. A resident worker loads the model once on the AMD GPU, then processes each image, avoiding a full model reload on every call. The project is packaged in Docker using the challenge-mandated ROCm PyTorch base image. It was validated on an AMD Instinct GPU for the required image formats, startup and inference limits, VRAM use, and a 298-image evaluation set. The larger evaluation achieved exact matches on all 82 samples with available ground-truth labels.