Embossed Metal OCR Model

I need a Python-based deep-learning OCR pipeline that can read raised or stamped characters on metal parts, even when the surface is rusty or dirty. The images will come as standard photos of the metal surfaces, not neat scans, so the model must handle glare, corrosion and inconsistent lighting.

A conventional Tesseract pass is not enough; I want a custom convolutional- or transformer-style network (PyTorch or TensorFlow/Keras) trained or fine-tuned specifically for embossed alphanumerics. If you already have techniques such as data-augmentation for metallic textures, domain adaptation or image-to-normal-map preprocessing in your toolkit, that will help accelerate the project.

Deliverables
• Clean, commented Python code that loads an image, locates the embossed region and returns the recognised string.
• A pretrained weight file (or training notebook) so I can reproduce results on new data.
• Brief README covering environment setup, dependencies and a quick demo command.

Acceptance will be based on the model achieving reliable character accuracy on a validation set of my own rusty-surface photos.