PP-OCRv5 mobile vs server, tested
1,904
of 1,955 words: mobile, English model, ONNX
Server lost to mobile on book pages
Two scanned English book pages, a Mac CPU, 2 threads a run.
| Models | Words | Time |
|---|---|---|
| Mobile, English model, PaddleOCR | 360 | 41 s |
| Mobile, English model, RapidOCR | 358 | 8 s |
| Mobile, RapidOCR | 350 | 9 s |
| Mobile, PaddleOCR | 349 | 45 s |
| Server, PaddleOCR, 960 px cap | 345 | 39 s |
| Server, RapidOCR | 341 | 104 s |
| Server, PaddleOCR | 337 | 129 s |
PaddleOCR's own recognition table agrees on English: mobile 66.00, server 64.70.
PaddleOCR server took 36 GB by default
PaddleOCR's default is type min with
limit_side_len 64. Type min only scales small pages up.
So detection got our 1920 px pages at full size.
There the server models peaked at 36 GB of memory and took 129 s a page. A cap on the long side fixed most of it:
text_det_limit_type="max", text_det_limit_side_len=960
With the cap: 7.3 GB, 39 s a page, and 8 more words.
The English model read 1,904 of 1,955
All 10 pages, mobile only. PaddleOCR server would have taken 21 minutes.
| Setup | Words | Time |
|---|---|---|
| v5 mobile, English model, RapidOCR | 1,904 | 8 s |
| PP-OCRv6 small, RapidOCR default | 1,895 | 4.5 s |
| Our Actor's code | 1,879 | 3.5 s |
| Tesseract 5.5.4 | 1,878 | 3.1 s |
| v5 mobile, RapidOCR | 1,869 | 8 s |
| v5 mobile, PaddleOCR | 1,856 | 47 s |
Screenshots: the English model won again
| Setup | Words right |
|---|---|
| v5 mobile, English model, RapidOCR | 97.5% |
| Apple Vision | 96.9% |
| v5 mobile, PaddleOCR | 94.9% |
| Our Actor's code | 94.4% |
| v5 mobile, RapidOCR | 93.2% |
| v5 server, RapidOCR | 76.0% |
In RapidOCR, the server models glued English words, like
youhavetobein theoficeorat. PaddleOCR's server run
read that line right.
RapidOCR ran mobile 6 times faster
- Same mobile models and pages: RapidOCR took 8 s a page, PaddleOCR 47 s. Part of the gap is size: RapidOCR first shrinks a page to 2,000 px on the long side, and PaddleOCR reads it at full size.
- RapidOCR's server detector cut a page into 203 boxes, most of them one word. PaddleOCR's made 29 lines.
- RapidOCR feeds its detectors mean and std 0.5, but these model files ask for ImageNet values.
- With ImageNet values, the page gave 65 boxes in 27 s, and 2 fewer words.
Pick mobile by name in PaddleOCR
from paddleocr import PaddleOCR
ocr = PaddleOCR(
text_detection_model_name="PP-OCRv5_mobile_det",
text_recognition_model_name="en_PP-OCRv5_mobile_rec",
use_doc_orientation_classify=False,
use_doc_unwarping=False,
use_textline_orientation=False,
)
for res in ocr.predict("page.jpg"):
print("\n".join(res["rec_texts"]))
lang="en", ocr_version="PP-OCRv5" gives you the server
detector with the English model. No language setting picks the mobile
detector.
Or run the ONNX files in RapidOCR
pip install rapidocr onnxruntime
from rapidocr import LangRec, ModelType, OCRVersion, RapidOCR
engine = RapidOCR(
params={
"Det.ocr_version": OCRVersion.PPOCRV5,
"Det.model_type": ModelType.MOBILE,
"Rec.ocr_version": OCRVersion.PPOCRV5,
"Rec.model_type": ModelType.MOBILE,
"Rec.lang_type": LangRec.EN,
}
)
print("\n".join(engine("page.jpg").txts))
On our first page, both scripts gave the same text as the runs above. RapidOCR has no English server model.
Five models and their sizes
| Model | Paddle | ONNX |
|---|---|---|
PP-OCRv5_mobile_det |
4.9 MB | 4.8 MB |
PP-OCRv5_server_det |
88 MB | 88 MB |
PP-OCRv5_mobile_rec |
17 MB | 17 MB |
PP-OCRv5_server_rec |
85 MB | 85 MB |
en_PP-OCRv5_mobile_rec |
8.0 MB | 7.9 MB |
All five are Apache-2.0. The two main recognizers read Chinese, Traditional Chinese, English and Japanese. The English one reads English only.
How we ran it
- A Mac with Apple M5 Max, macOS 26.6. PaddleOCR 3.7.0 with paddlepaddle 3.3.1, and RapidOCR 3.10.0 with onnxruntime 1.32.0. Most runs: two at a time, 2 threads each, at low priority.
- The scan: 10 proofread pages from two public domain books on Wikisource, 1920 px JPEGs, 1,955 words. The 381-word table uses its first 2 pages.
- PaddleOCR ran with its three page-fix models off. RapidOCR ran with its defaults, which include its line-turn model and a 2,000 px cap on the long side, plus the v5 model names.
- Score: true words found in order. Screenshots: the 16 at 1x from our screenshot test, words right in any order.
- Times are rough: the Mac ran other work too. The Tesseract, Actor and Apple Vision rows are from our earlier runs today.
- I stopped the PaddleOCR server run on screenshots after 6 of 16 shots, to keep the daytime test short. RapidOCR server with the 960 px cap found 342 words in 109 s a page.
Scanned pages to text, no install
Run it yourself. Our Apify Actor turns a scanned PDF into Markdown or text. $0.006 a scanned page. Its code found 1,879 of these 1,955 words.
Or send us the scan. Tack, our AI, does the job for $29, up to 50 pages. First page free.
Questions people ask
Is PP-OCRv5 server better than mobile?
Not on our English pages. Mobile found more words on book scans and on screenshots, and it was faster. PaddleOCR's own tables favor server for Chinese and handwriting.
Which PP-OCRv5 model is best for English?
Of the pairs we tested, en_PP-OCRv5_mobile_rec with PP-OCRv5_mobile_det. In RapidOCR it found 1,904 of 1,955 words on our scan and 97.5% of the words on 16 screenshots.
Where do I get PP-OCRv5 ONNX files?
RapidOCR downloads them for you from ModelScope on first use. PaddlePaddle also posts ONNX copies on Hugging Face, such as PaddlePaddle/PP-OCRv5_mobile_det_onnx. Both are Apache-2.0.
Is PP-OCRv5 still the default?
No. PaddleOCR 3.7.0 uses PP-OCRv6 medium by default, and RapidOCR 3.10.0 uses PP-OCRv6 small. Both still run PP-OCRv5 when you name it.
Does your Actor use PP-OCRv5?
No. It reads scans with Tesseract. Its second reader, for English pages that Tesseract reads badly, is RapidOCR with PP-OCRv6 small.