Scanned books, read in page order

Why a readable scan can produce scrambled text

Text detection and reading order are different jobs. A detector can recognise each line and still return lines from both facing pages at the same height. Joining that output makes the left and right pages alternate. That is the failure documented in the RapidOCR vs Tesseract write-up.

This is a schematic, not OCR output. The home page's scan exhibit shows a separate Darwin spread; it is an illustration, not the benchmark corpus.

Find the gutter before joining text

A fixed midpoint only works when the gutter is actually centred. Inspect the scan before assuming the middle of its image is a safe cut. The documented B.L.A.S.T. approach looks for a vertical white gap in the central band of a landscape spread, then handles each page separately. It does not establish a general curved-seam detector.

The blog describes a spread test of width greater than 1.1 times height and a search in the central 40 to 60% of the width. Those are pipeline heuristics, not rules for every book. A dark gutter, an off-centre scan or text close to the binding can defeat a simple gap search.

Illustrative Python sketch, not B.L.A.S.T. source

This runnable OpenCV/numpy sketch finds a straight low-ink column in that central band and writes two images. It does not perform OCR, detect columns or reproduce the benchmark pipeline. Install opencv-python and numpy, save as split_spread.py, then run python split_spread.py spread.png.

import argparse
from pathlib import Path

import cv2
import numpy as np

parser = argparse.ArgumentParser()
parser.add_argument("image", type=Path)
args = parser.parse_args()
image = cv2.imread(str(args.image))
if image is None:
    raise SystemExit("Could not read the input image")

height, width = image.shape[:2]
if width < 10 or width <= 1.1 * height:
    raise SystemExit("Expected a landscape book spread; inspect it first")

gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
_, ink = cv2.threshold(
    gray, 0, 255, cv2.THRESH_BINARY_INV | cv2.THRESH_OTSU
)
column_ink = ink.mean(axis=0)
smoothed = np.convolve(column_ink, np.ones(9) / 9, mode="same")
start, stop = int(0.4 * width), int(0.6 * width)
split = start + int(np.argmin(smoothed[start:stop]))

for side, crop in (("left", image[:, :split]),
                   ("right", image[:, split:])):
    output = args.image.with_name(f"{args.image.stem}-{side}.png")
    if not cv2.imwrite(str(output), crop):
        raise SystemExit(f"Could not write {output}")
    print(output)

Review both crops before a batch run. A minimum in an ink profile is only a candidate split; this sketch provides no confidence check and may choose blank page space instead of the gutter. Keep the original image so you can revise the split without losing text.

Preserve columns and reading order inside each page

For left-to-right text, finish the left page before starting the right page. Within each page, identify columns before sorting lines top to bottom. Sorting every text box by its vertical coordinate alone recreates the same interleaving problem at the column level.

The documented pipeline finds vertical gaps, groups boxes into lines using glyph-scaled tolerances, sorts within lines, and joins nearby lines into blocks. That is the level supported by the write-up and B.L.A.S.T. case study; use the pinned OCR repository for the implementation. The sketch above stops before these stages.

On this corpus, mean reading-order tau was 0.668 for raw RapidOCR,0.993 for B.L.A.S.T. and 0.995 for raw Tesseract. Null scores are excluded from those means. Use the benchmark tables and methodology to see the valid page counts and scoring limits.

When Tesseract is the better starting point

Start with Tesseract for printed prose when its automatic segmentation already reads the pages correctly. In the published comparison it won 9 of 14 pages against B.L.A.S.T. and had lower mean WER: 0.279 versus 0.449. It averaged 3.7 seconds per page against 16.5, a 4.5× speed advantage on the measured OCR calls.

Those times exclude restoration. They are not full document processing times. Adding a layout stage should solve an observed failure on your sample; it should not be a requirement imposed on pages that already read correctly.

When layout handling earns the extra work

Add explicit layout handling when inspection shows facing pages or columns being mixed. In this benchmark B.L.A.S.T. beat raw Tesseract on the contents page, some endnotes and the sparse pages. The sparse-page effect is why its lower mean CER needs the per-page results alongside it.

Correct order still leaves word cleanup. The comparison documents missing spaces after punctuation, joined words, line-end hyphens and running headers. Review these before sending text into search or retrieval. This benchmark does not establish performance on tables or handwriting; evaluate those separately if your documents contain them.

Pre-flight checks before processing a book

Frequently asked questions