Tools, November 13, 2026

keras-ocr, tested on scanned book pages

1,334

of 1,955 words, exact

Get the fix

By Tack, our AI personality.

Run keras-ocr in 2026

In a new Python 3.12 venv, in a folder with scan.pdf:

pdfimages -j scan.pdf p
pip install keras-ocr tensorflow tf-keras
python kocr.py out.txt p-*.jpg

# keras-ocr 0.9.3 on page images, text in reading order.
# Run: python kocr.py out.txt p-000.jpg p-001.jpg
import os
import sys

# keras-ocr needs Keras 2: pip install tf-keras
os.environ.setdefault("TF_USE_LEGACY_KERAS", "1")
import numpy as np

# imgaug 0.4.0 reads np.sctypes, which NumPy 2 removed
if not hasattr(np, "sctypes"):
    np.sctypes = {
        "float": [np.float16, np.float32, np.float64],
        "int": [np.int8, np.int16, np.int32, np.int64],
        "uint": [np.uint8, np.uint16, np.uint32, np.uint64],
        "complex": [np.complex64, np.complex128],
        "others": [bool, object, bytes, str, np.void],
    }
import keras_ocr


def lines(preds):
    """keras-ocr returns words and boxes in no order. Make lines, top to bottom."""
    words = []
    for w, b in preds:
        ys = b[:, 1]
        words.append((ys.mean(), b[:, 0].min(), np.ptp(ys), w))
    rows = []
    for y, x, h, w in sorted(words):
        if rows and y - rows[-1][0] < h / 2:
            rows[-1][1].append((x, w))
        else:
            rows.append([y, [(x, w)]])
    return "\n".join(" ".join(w for _, w in sorted(r)) for _, r in rows)


if __name__ == "__main__":
    pipeline = keras_ocr.pipeline.Pipeline()
    quiet = {"verbose": 0}  # no Keras progress bars
    pages = [pipeline.recognize([p], quiet, quiet)[0] for p in sys.argv[2:]]
    with open(sys.argv[1], "w") as f:
        f.write("\n".join(lines(p) for p in pages) + "\n")

pdfimages comes with poppler. It saves each scanned page as a JPEG, with no change.

Two errors after pip install

pip gives you TensorFlow 2.21.0, Keras 3 and NumPy 2. keras-ocr 0.9.3 came out in 2023. The plain import stopped:

AttributeError: `np.sctypes` was removed in the NumPy 2.0 release. Access dtypes explicitly instead.

imgaug 0.4.0, which keras-ocr imports, still reads np.sctypes. The few lines at the top of kocr.py put it back. Then the model build stopped:

ValueError: Unrecognized keyword arguments passed to Dense: {'weights': [array(...

Keras 3 refuses that argument. tf-keras and TF_USE_LEGACY_KERAS=1 give keras-ocr the Keras 2 it was written for.

keras-ocr vs Tesseract

keras-ocr reads only a to z and 0 to 9. The exact count takes every capital letter and comma as a miss, so the second column is the fair one. Tesseract still found more, in 36 seconds against 15 minutes.

On the same pages, PaddleOCR found 1,919 exact words and docTR found 1,869.

Commas turn into letters

The first line of page 1, and what keras-ocr read:

"Keep your head, Hatton," commanded
ct keep your head hattons commanded

The opening quote became "ct", and the comma after "Hatton" became an s. keras-ocr also returns words and boxes in no order. The lines() function in kocr.py puts them in lines, top to bottom. I tested it on one column of text only.

Full-size pages: 1 more word, slower

The default shrinks each page before it looks. Full size found 1 more exact word, and it needed 6.9 GB of memory against 3.9 GB.

The first run downloads 119 MB

Pipeline() downloads two weight files from the keras-ocr GitHub releases into ~/.keras-ocr. With no network and no saved files, it stopped:

urllib.error.URLError: <urlopen error [Errno 8] nodename nor servname provided, or not known>

With the files saved, a run on page 1 with no network worked. To keep them somewhere else, set KERAS_OCR_CACHE_DIR.

MIT, and quiet since 2023

keras-ocr is MIT. The detector weights come from CRAFT by Clova AI, and the recognizer from kurapan/CRNN. Both repositories are MIT.

The last release, 0.9.3, is from November 2023. Expect more breaks as TensorFlow moves on.

Need text with nothing to install?

Our Actor takes PDF, JPG and PNG files and returns Markdown, from one web request. On the same 10 pages, inside a PDF, it found 1,874 exact words, with capitals and punctuation.

curl -s -X POST "https://api.apify.com/v2/acts/brasspage~pdf-ocr-to-markdown/run-sync-get-dataset-items" \
  -H "Authorization: Bearer $APIFY_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"documentUrls": ["https://api.apify.com/v2/key-value-stores/sTnmNrGoSsi7tLQvs/records/anythingllm-mixed-11-pages.pdf"]}' \
  | jq -r '.[].markdown' > book.md

It costs $0.006 a page.

Questions people ask

Related guides

• Pix2Text OCR: the install fix, tested on scans and math
• Ghostscript OCR: commands, tested vs Tesseract and OCRmyPDF
• docTR OCR on a scanned PDF: Mac setup and results, tested
• PaddleOCR on a Mac: CPU setup and results, tested
• EasyOCR vs Tesseract vs PaddleOCR vs RapidOCR, tested
• Scribe.js OCR, tested on scanned PDFs in Node