Solving NDLOCR Column Layout Recognition with Histogram Analysis

Goal

OCR processing of a vertical Japanese book (4-column layout)

Environment

Item	Details
OS	Windows 11
Python	3.14.0

---

Attempt 1: NDLOCR Alone

Result: Partially works but layout recognition is weak

• Column detection fails, breaking reading order
• NDLOCR’s layout analysis was designed for older vertical-text books and struggles with modern multi-column layouts

---

Attempt 2: Layout Parser + NDLOCR

Result: Installation failure

• Layout Parser itself installs fine with pip install layoutparser
• Detectron2 build error on Windows
  • pip install "detectron2@git+..." → build failure
  • C++ compiler issues in Windows environment

---

Attempt 3: PyMuPDF + Histogram Analysis (Alternative)

Result: Works

• No Detectron2 needed
• Vertical histogram detects column boundaries (valleys)
• Fixed 4-column split → pass to NDLOCR

---

Final Setup

PDF
 ↓
extract_blocks.py (PyMuPDF + histogram analysis)
 ↓
4 split images (page01_1.jpg ~ page01_4.jpg)
 ↓
NDLOCR (OCR only)
 ↓
Text output

Requirements

• Python + PyMuPDF, PIL, numpy
• Docker + NDLOCR
• extract_blocks.py (histogram analysis script)

---

extract_blocks.py

import fitz  # PyMuPDF
from PIL import Image
import io
import os
import numpy as np

# --- Configuration ---
pdf_path = r"C:\ndlocr_work\input\test.pdf"
output_dir = r"C:\ndlocr_work\input\img"
# ---------------------

os.makedirs(output_dir, exist_ok=True)

try:
    pdf_document = fitz.open(pdf_path)
except fitz.errors.FitzError as e:
    print(f"Error: Cannot open PDF file. Check the path: {e}")
    exit()

print(f"Starting to process {len(pdf_document)} pages.")

def find_split_point(histogram, center_y, search_range):
    """Find the center of a valley in the histogram"""
    start = max(0, center_y - search_range)
    end = min(len(histogram), center_y + search_range)
    
    # Extract histogram segment within search range
    segment = histogram[start:end]
    
    # Find the valley depth (minimum value)
    min_val = np.min(segment)
    
    # Get all indices of the minimum value (valley bottom)
    valley_indices = np.where(segment == min_val)[0]
    
    # Get the center index of the valley bottom
    middle_of_valley = valley_indices[len(valley_indices) // 2]
    
    # Return coordinate in the full histogram
    return start + middle_of_valley

for page_num in range(len(pdf_document)):
    page = pdf_document[page_num]
    print(f"\n--- Page {page_num + 1} ---")

    # 1. Render page at high resolution
    mat = fitz.Matrix(3, 3)
    pix = page.get_pixmap(matrix=mat, alpha=False)
    img = Image.open(io.BytesIO(pix.tobytes("jpeg")))

    # 2. Binarize
    gray_img = img.convert('L')
    binary_img = gray_img.point(lambda x: 0 if x < 200 else 255, '1')

    # 3. Calculate vertical histogram
    np_img = np.array(binary_img)
    histogram = np.sum(255 - np_img, axis=1)

    # 4. Detect 3 split points at valley centers
    height, width = np_img.shape
    search_range = int(height * 0.05)  # Search within ±5% of center
    
    split_points = [
        find_split_point(histogram, int(height * 0.25), search_range),
        find_split_point(histogram, int(height * 0.50), search_range),
        find_split_point(histogram, int(height * 0.75), search_range)
    ]
    
    split_points.sort()
    print(f"Detected split point Y coordinates: {split_points}")

    # 5. Crop and save image segments
    boundaries = [0] + split_points + [height]

    for i in range(4):
        y_start = boundaries[i]
        y_end = boundaries[i+1]
        
        segment = img.crop((0, y_start, width, y_end))
        output_path = os.path.join(output_dir, f"page{page_num+1:02d}_{i+1}.jpg")
        segment.save(output_path, "JPEG", quality=95)
        print(f"  Column {i+1} saved: {output_path}")

pdf_document.close()
print("\nAll processing complete!")