vladlinv/ru-ocr-benchmark-hard

A benchmark for evaluating OCR in Russian, consisting of two parts: pages and crops. The pages part contains 155 images covering 31 source documents in five variants: originals, scans, low-quality images, print defects, and distortions. Reference text and coordinates for each character enable evaluation of complete OCR pipelines. The crops part contains 3,182 ready-made text crops for evaluating text recognition independently of detection. Both parts include challenging examples with varied fonts, image quality, and content.

Note: In future updates, the benchmark will be expanded with additional challenging examples and handwritten text samples.

Benchmark Results

Results on ru-ocr-benchmark-hard, a challenging benchmark covering distortions, print defects, and low-quality images.

Full pages — complete OCR pipeline (155 images)

Crops — text recognition (3182 images)

Model 1-NED SER WER Accuracy Yandex Vision OCR 0.967687 20.21% 6.74% 18.06% PP-OCRv6 Medium RU (For RapidOCR) DET: PP-OCRv6 Medium 0.977351 16.14% 5.59% 13.55% PP-OCRv6 Tiny RU (For RapidOCR) DET: PP-OCRv6 Small 0.971663 27.19% 9.04% 3.87% Occular-OCR (SVTR-T) 0.958628 42.07% 12.34% 0.00% PP-OCRv5 Cyrillic (RapidOCR) DET: PP-OCRv5 Server 0.841356 50.28% 27.10% 0.00% PP-OCRv5 ESlav (RapidOCR) DET: PP-OCRv5 Server 0.833810 57.50% 30.78% 0.00% Tesseract 5 Best 0.761598 61.68% 36.37% 0.00%

Model 1-NED SER WER Accuracy PP-OCRv6 Medium RU (For RapidOCR) 0.951630 28.13% 16.02% 71.87% Yandex Vision OCR 0.845650 39.25% 22.29% 60.75% PP-OCRv6 Tiny RU (For RapidOCR) 0.908217 48.18% 27.71% 51.82% Occular-OCR (SVTR-T) 0.884307 50.97% 29.24% 49.03% PP-OCRv5 Cyrillic (RapidOCR) 0.803864 67.13% 46.66% 32.87% PP-OCRv5 ESlav (RapidOCR) 0.790004 69.58% 50.21% 30.42% Tesseract 5 Best 0.659237 78.13% 60.65% 21.87%

---

Detailed metrics by category

Distortions (31 pages)

Model	1-NED	SER	WER	Accuracy
PP-OCRv6 Medium RU + UVDoc	0.964851	17.66%	9.12%	9.68%
PP-OCRv6 Tiny RU + UVDoc	0.973516	34.05%	12.49%	0.00%
Occular-OCR (SVTR-T)	0.952869	41.54%	13.56%	0.00%
Yandex Vision OCR	0.900997	40.81%	18.80%	6.45%
PP-OCRv5 Cyrillic (RapidOCR) + UVDoc	0.843278	50.82%	25.02%	0.00%
PP-OCRv5 ESlav (RapidOCR) + UVDoc	0.837129	58.43%	28.46%	0.00%
Tesseract 5 Best	0.602607	85.89%	59.98%	0.00%

Originals (31 pages)

Model	1-NED	SER	WER	Accuracy
Yandex Vision OCR	0.995442	6.34%	1.25%	35.48%
PP-OCRv6 Medium RU	0.990530	7.27%	2.04%	22.58%
PP-OCRv6 Tiny RU	0.966852	11.40%	5.39%	3.23%
Occular-OCR (SVTR-T)	0.983030	30.20%	6.65%	0.00%
Tesseract 5 Best	0.963236	25.94%	7.23%	0.00%
PP-OCRv5 Cyrillic (RapidOCR)	0.838570	28.85%	21.08%	0.00%
PP-OCRv5 ESlav (RapidOCR)	0.833360	36.49%	23.24%	0.00%

Scans (31 pages)

Model	1-NED	SER	WER	Accuracy
Yandex Vision OCR	0.993804	9.09%	1.92%	22.58%
PP-OCRv6 Medium RU	0.988223	9.76%	3.00%	22.58%
PP-OCRv6 Tiny RU	0.990307	17.89%	4.72%	9.68%
Occular-OCR (SVTR-T)	0.975689	36.58%	9.96%	0.00%
Tesseract 5 Best	0.916912	42.79%	15.43%	0.00%
PP-OCRv5 Cyrillic (RapidOCR)	0.848345	44.91%	24.21%	0.00%
PP-OCRv5 ESlav (RapidOCR)	0.840750	53.67%	28.07%	0.00%

Low quality (31 pages)

Model	1-NED	SER	WER	Accuracy
Yandex Vision OCR	0.974547	17.21%	4.69%	25.81%
PP-OCRv6 Medium RU	0.974096	19.52%	6.39%	9.68%
PP-OCRv6 Tiny RU	0.969358	30.51%	10.30%	6.45%
Occular-OCR (SVTR-T)	0.957408	40.91%	12.44%	0.00%
Tesseract 5 Best	0.877041	55.26%	23.97%	0.00%
PP-OCRv5 Cyrillic (RapidOCR)	0.851886	53.10%	28.19%	0.00%
PP-OCRv5 ESlav (RapidOCR)	0.840663	59.73%	31.89%	0.00%

Print defects (31 pages)

Model	1-NED	SER	WER	Accuracy
Yandex Vision OCR	0.973643	27.55%	7.02%	0.00%
PP-OCRv6 Medium RU	0.969052	26.39%	7.41%	3.23%
PP-OCRv6 Tiny RU	0.958282	41.88%	12.32%	0.00%
Occular-OCR (SVTR-T)	0.924145	59.53%	19.08%	0.00%
PP-OCRv5 Cyrillic (RapidOCR)	0.824699	73.38%	36.99%	0.00%
PP-OCRv5 ESlav (RapidOCR)	0.817147	79.01%	42.24%	0.00%
Tesseract 5 Best	0.448197	96.06%	75.24%	0.00%

Links