Offline Video Number Recognition AI

I have a folder of MP4 recordings that contain all kinds of on-screen numbers—timestamps, counters, device read-outs, you name it. I need an offline application (source code is fine) that can automatically detect, read, and export every numerical value it finds as the video plays frame by frame. Nothing should leave the machine, so the entire pipeline must run locally without any cloud calls.

Key details you can rely on: the videos are pre-recorded (no live streams) and always in MP4 format; the numbers are generic, not limited to scores or license plates. I want an AI-driven approach so that, after the first delivery, I can feed the model corrections or new examples and watch its accuracy climb over time. Think of a small feedback interface or a retraining script that lets me tweak performance without rewriting the whole program.

I am comfortable with mainstream computer-vision stacks such as OpenCV, TensorFlow, or PyTorch, so feel free to build on those. The final package should include:

• A runnable app or script that ingests an MP4 file, detects every distinct numerical instance, and outputs the readings (JSON or CSV is fine).
• A simple mechanism—command-line flag, YAML config, or lightweight GUI—that lets me supply corrected readings and trigger incremental training.
• Clear setup instructions plus a brief note on how the model can be extended to new video resolutions or number fonts.

Acceptance criteria: on a short sample video I will provide, the tool should correctly capture at least 90 % of the visible numbers on first run, store timecodes alongside each reading, and retrain successfully when I give it a handful of edits.

If that matches your expertise, I’m ready to test your prototype and iterate quickly.