High-Accuracy Python OCR Engine
Budget: $10 – $1,000 USD
I’m sitting on a bank of roughly one million black-and-white images, each a tight 240 × 30 pixels and expected to hold either six or eight digits. The catch is that the digits drift a little in size and position, the backgrounds are less than pristine, and every “obvious” off-the-shelf AI solution I have tried tops out at about 93 % accuracy. I need 99 %.
attached example of pictures
500'000 pictures wil be given to selected candidat
this is not for novice!
on offer please write 500'000 so i know you read spec...
What I’m after is a clean Python program that ingests a single image and returns the full number it contains. From experience, tackling the task digit-by-digit rather than in one sweep gives the best shot, so please architect your approach with that in mind.
You will have access to a sizeable, fully-labeled training set, so there’s no need to build annotations from scratch. Feel free to bring in PyTorch, TensorFlow, Keras, OpenCV or any other Python-friendly tools you deem helpful, provided the final pipeline still meets the 99 % bar on my held-out test batch.
When you deliver, I expect:
• Source code with clear setup instructions
• A trained model (or reproducible training script)
• A short validation report demonstrating ≥99 % digit-level accuracy on unseen data
That’s it—simple to state, challenging to achieve. If you’re confident you can push past that last six percent, let’s talk.
attached example of pictures
500'000 pictures wil be given to selected candidat
this is not for novice!
on offer please write 500'000 so i know you read spec...
What I’m after is a clean Python program that ingests a single image and returns the full number it contains. From experience, tackling the task digit-by-digit rather than in one sweep gives the best shot, so please architect your approach with that in mind.
You will have access to a sizeable, fully-labeled training set, so there’s no need to build annotations from scratch. Feel free to bring in PyTorch, TensorFlow, Keras, OpenCV or any other Python-friendly tools you deem helpful, provided the final pipeline still meets the 99 % bar on my held-out test batch.
When you deliver, I expect:
• Source code with clear setup instructions
• A trained model (or reproducible training script)
• A short validation report demonstrating ≥99 % digit-level accuracy on unseen data
That’s it—simple to state, challenging to achieve. If you’re confident you can push past that last six percent, let’s talk.