OCR benchmarks
Reading text from scanned pagesText recovery and reading order on 32 English pages: scans, forms, columns and small print.
Overview
Capability & performance
Text checks passed ↑ · time and size ↓Benchmark breakdown
Eight pages in each categoryCompare models
— means no verified measurement on this device. Page timing includes image loading, preprocessing, text detection and recognition. Model loading and the two warm-up pages are measured separately.
Missed text
Examples from the saved answersTest details
Fixed pages, original model settingsBenchmark progress
Scores compare recovered passages with the expected text. Speed and memory are measured on the Pi 5.
Other OCR models
Researched candidates. A listing here does not mean the model has been benchmarked or runs on the Pi.
Results table · 7 configurations
Text recovery and reading order on 32 English document pages, scored with 72 checks. The overall score weights four document categories equally. Higher is better. This is separate from table, equation and page-layout quality. A dash means no matching measurement is available.
| Model | Text checks passed · % ↑ | Pi 5 seconds per page ↓ | Peak memory · MiB ↓ |
|---|---|---|---|
| PP-OCRv4 mobile | 31 | 29.5 | 688 |
| PP-OCRv6 tiny | 31 | 5.3 | 576 |
| PP-OCRv6 small | 32 | 18.6 | 775 |
| PP-OCRv6 medium | 33 | 78.7 | 1,617 |
| PP-OCRv5 mobile | 30 | 24.9 | 766 |
| Indra OCR 1.1 | 60 | 2 | 710 |
| PP-OCRv5 mobile · AI HAT | 11 | 2.8 | 187 |