AMD VisionOCR: Robust OCR on ROCm

AMD VisionOCR is a GPU-accelerated OCR application built for the Lablab x AMD AI Academy Mini Challenge 2. It uses Qwen2.5-VL-3B-Instruct with AMD ROCm to extract text from challenging real-world images, including license plates, stop signs, speed-limit signs, advisory speed plaques, and road-work warning signs. The OCR prompt is designed around the challenge requirements, including preserving relevant regional characters, excluding license-plate slogans and state names, reading multi-line signs in order, and returning advisory speed numbers only. The application supports PNG, JPEG, and TIFF inputs and handles image orientation before inference. The model and application are packaged into a self-contained Docker image using the mandated ROCm PyTorch base image, with the model weights included locally. Testing achieved 10/10 on the prepared OCR benchmark, followed by additional stress testing involving blur, darkness, low contrast, rotation, glare, and different image sizes. The final container is publicly available through GitHub Container Registry.