Price Recognition in Retail with PaddleOCR

Danil Muzafarov24 May 2025

Problem

A retailer that prices dynamically needs to know what neighboring stores charge. Employees photograph a product on a competitor's shelf, a detector finds the price, and an OCR model reads it into the price field of the result. The price goes into a database used by pricing algorithms, so every character from rubles to kopecks must be right. The production recognizer was an attention-based model on TensorFlow. The team wanted a replacement that is easy to fine-tune, runs on OpenVINO and is no worse than production in accuracy and speed.

Example from the article: a product costs 59 rub 99 kop in store 1 and 55 rub 99 kop in store 2. The price crop reads "269" with kopecks "90" and must produce price: "26990".

Solution in brief

Train the recognition-only algorithms of PaddleOCR on a small price dataset and keep those that meet the requirements. Fine-tune the finalists on about 1M samples and compare them with production on 10,000 real monitorings. PP-OCRv3 beat production by about 0.3% and made the pipeline faster.

Questions this case answers
  • How to read short numeric values exactly from cropped images with a replacement for a production recognizer that must be at least as accurate and faster
  • How to choose an OCR framework to fine-tune on your own crops when the result must run on an optimized inference runtime
  • How to read competitor prices from shelf photos for dynamic pricing
  • How to replace a production OCR model with a faster open-source one without losing accuracy

Tools and methods

  • PaddleOCR — open-source OCR framework with about 14 text-recognition algorithms, pretrained weights and configs
  • PP-OCRv3, Rosetta, SVTR, RobustScanner — recognizers compared on price crops
  • Exact-match accuracy — the main metric, tied to the business use of the price
  • OpenVINO (via ONNX) — inference runtime of the existing pipeline
  • ClearML — experiment tracking
Continue reading — it's free

Google or a one-time email code. No password.