AMD OCR Vision Challenge

This project is an OCR solution developed for the Lablab x AMD AI Academy Challenge – Mini Challenge 2. It uses Qwen2-VL-7B-Instruct, a vision-language model, to recognize text from real-world images such as US license plates, Chinese license plates, stop signs, speed limit signs, work zone signs, and advisory speed plaques. The solution is designed to handle different image conditions, including blur, low light, glare, noise, and off-axis views. A task-specific OCR prompt guides the model to extract only the text belonging to the target object, preserve the correct reading order, and avoid unnecessary explanations or surrounding text. The model runs on AMD ROCm and is packaged in a Docker container based on the mandated AMD ROCm PyTorch base image. The container loads the model once and processes the input image through the required command-line interface, producing a JSON output containing the recognized text. The solution was tested on the provided challenge examples, including US and Chinese license plates, traffic signs, and challenging image conditions. The Docker image was successfully built, tested with an AMD Instinct MI300X GPU, and published as a public Docker Hub image for evaluation.