I need to add a reliable mask generator to my ANPR pipeline that pinpoints the license-plate region in still JPEG photographs. The task is limited to detection—no character recognition for now—so the model simply has to return an accurate bounding box or, better, a pixel-level mask for every input image. Please base the solution on an R-CNN family network (Mask R-CNN, Faster R-CNN, or a comparable variant) written in Python. The script should accept a folder of JPEGs, run inference, and produce either a JSON/CSV file with bounding-box coordinates or a set of alpha-masked images. I will judge success on detection accuracy: roughly 95 % recall across my 2 000-image test set that mixes daylight and night shots. You are free to start from public weights—COCO or similar—so long as you document all datasets and training steps. Reproducibility is key; I want to be able to retrain the model from scratch on my side. Deliverables (to be accepted): • Clean, well-commented Python code and requirements.txt • Trained weights (.pth, .ckpt, or .h5) • CLI script plus a minimal REST endpoint for batch inference • README explaining data prep, training, and inference commands OpenCV, PyTorch, TensorFlow, or Keras are all acceptable; pick whichever you are most comfortable with, and keep the dependency list lean.