Is it possible to train and run a small AI model on a Raspberry Pi 5 to solve text-based CAPTCHAs?

Question

I’m trying to understand whether it’s actually feasible to train and run a small AI model on a Raspberry Pi 5 (16 GB RAM) that can solve simple text-based CAPTCHAs, the kind that contain a few letters or numbers with mild distortion, rotation, or background noise.

I’m not looking for an existing or pre-trained model. The question is whether a small custom model could realistically be trained to perform that task with good accuracy. For example, could a model be trained on synthetic CAPTCHA images, achieve reliable results when recognizing characters?

The goal would be to run everything fully on-device, without relying on external APIs or cloud inference. I’m mainly wondering whether this is technically possible, and if so, what sort of architectures (e.g. compact CNNs, MobileNet-like models, or simple OCR-style networks) might make it achievable within the limits of the Raspberry Pi 5’s CPU or GPU.

This question is purely for educational and research purposes, I’m interested in exploring the limits of small models for OCR-like tasks, not in breaking or bypassing any real CAPTCHA systems.

Any insights into whether such a small model could be trained successfully, and what factors would determine its feasibility, would be appreciated.

Answer 1

that is a question better asked on https://stats.stackexchange.com/