Offline Document Asseessbility (PDF)

Document accessibility ensures equal access to information for everyone, regardless of their abilities. It not only helps foster a more inclusive society but is also an ethical responsibility to respect diverse audience needs. Besides, investing in document accessibility can benefit businesses by expanding their audience and demonstrating a commitment to ethical practices and social responsibility.

Here is a Python CLI tool designed to detect and address WCAG accessibility issues identified in offline documents, targeting PDF. It aims to promote a user-friendly experience for everyone and enables users to access documents with assistive technology.

Installation

git clone https://github.com/caraaaaa/doc_accessibility.git
conda env create -f env/environment.yml
conda activate doc_accessibility_cli

Features

• Searchable PDF Creator
• Text Alternatives for Non-text Content
  • Identify Non-text Content
  • Generate Description for Non-text Content
    • Classify image-of-text and non-text-image
    • Image Captioning
    • OCR
• Text Representation
  • Text Constrast
  • Line Space

Searchable PDF Creator

Convert scanned PDF document to searchable format using TesseractOCR.

Criteria: User can read or extract the words using assistive technologies, or manipulate the PDF for accessibility.

WACG guideline: 1.4.5

Scanned PDF	Searchable PDF

Usage Instruction

python script/scanned2searchable.py [-o OUTPUT_PDF_PATH] [-s] input_pdf_path

Default output path: readable_pdf.pdf

optional arguments:
  -o OUTPUT_PDF_PATH, --output_pdf_path OUTPUT_PDF_PATH
                        The path for the output searchable PDF file.
  -s, --show_result     Show text of the searchable PDF after OCR

Text Alternatives for Non-text Content

Provide descriptions for images inside searchable PDF using image classification, transformer model and OCR.

Criteria: All non-text content (e.g. images, formulas) that is presented to the user has a text alternative that serves the equivalent purpose

WACG guideline: 1.1

Overview

flowchart LR
    A[Extract Images] -->|Optional| B[Save Images]
    A --> C[Check Alt Text]
    C --> D[Classification]
    D -->|image of text| F[Perform OCR]
    D -->|non-text image| G[Generate Image Caption]
    F --> H[Provide Alt Text Suggestions]
    G --> H
    H -->|Optional| I[Output PDF with Bounding Box]

Loading

Usage Instruction

python script/extract_PDF_image.py [--output_img] [--output_folder OUTPUT_FOLDER] [--draw_bbox] [--output_pdf_path OUTPUT_PDF_PATH] [--captioning] input_pdf_path

optional arguments:
  --output_img          Output images extracted from the PDF.
  --output_folder OUTPUT_FOLDER
                        The directory for the output images.
  --draw_bbox           Output PDF with bounding box on images.
  --output_pdf_path OUTPUT_PDF_PATH
                        The path for the output PDF file with bounding box.
  --captioning          Generate caption for images.

Default output folder: pdf_image Default output pdf path: image_bbox.pdf

Identify Non-text Content

Part of Text Alternatives for Non-text Content

Identify images which without alternative text using PyMuPDF and Pillow.

Usage Instruction

python script/extract_PDF_image.py --draw_bbox input_pdf_path

Default output path: image_bbox.pdf

Image of Text Classifier

Part of Text Alternatives for Non-text Content

Classifiy if an image (such as JPG or PNG) primarily contains text before performing OCR or image captioning.

Image of Text	Non-text image

Fine-tuned a image classification model

• Data: online-sourced images
• Model: ResNet
• Tools: fastai
• Training Script: Google Colab

Usage Instruction

python script/image_of_text.py [--show_score] input_pdf_path

Optional arguments:
  --show_score    Show the classification score

Image Captioning

Part of Text Alternatives for Non-text Content

Generate descriptive caption for non-text image using Transformer model.

Non-text image	Caption
	Indication of correct signature

Fine-tuned a Transformer model

• Data:

  • Non-text images extracted from PDF
  • Captioned Manually
  • Dataset card

• Model:

  • Pre-trained: GenerativeImage2Text
  • Fined-tuned: model card and inference API

• Tools: HuggingFace Transformer, PyTorch

• Training Script: Google Colab

Usage Instruction

python script/generate_caption.py <input_image_path>

OCR

Part of Text Alternatives for Non-text Content

Extract text from image-of-text using TesseractOCR

Usage Instruction

python script/image_of_text.py input_pdf_path

Text Constrast

Identify low contrast text inside PDF using image segmentation and contrast ratio analysis.

Criteria: The visual presentation of text and images of text has a contrast ratio of at least 4.5:1

WACG guideline: 1.4.3

Overview

flowchart LR
    A[Extract Text Bounding Box]
    A --> C[Image Segmentation] -->|Optional| B[Save model prediction]
    C --> D[Calculate contrast]
    D -->|Optional| E[Output PDF with Bounding Box]

Loading

Image of Text	Predicted Text Segmentation	Enought Constrast	Low Contrast

Fine-tuned a Transformer model

• Data:
  • Generate synthetic image of text with segmentation mask using Pillow
  • Dataset Card
• Model:
  • Pre-trained: SegFormer
  • Fined-tuned: model card
• Tools: HuggingFace Transformer, PyTorch
• Training Script: Google Colab

Calculate the contrast ratio

• Luminance (brightness) of the colors:

$$L = 0.2126\times R+0.7152\times G+0.0722\times B$$

$$\small\text{where R, G, and B are normalized to 0-1}$$

• Contrast Ratio $$\frac{L2+0.05}{L1+0.05}$$

$$\small\text{where L1 is the luminance of the lighter color, either text or background}$$

Usage Instruction

• Generate synthetic image of text

python script/synthetic_text_seg.py [--sample_no SAMPLE_NO] [--output_folder OUTPUT_FOLDER] [--font_folder FONT_FOLDER]

Default output folder: image_of_text

Default font input folder: font

• Identify low contrast text

python script/contrast_pdf.py [--output_bbox_img] [--output_dir OUTPUT_DIR] [--draw_bbox] [--output_pdf_path OUTPUT_PDF_PATH] [--bbox_extractor {PyMyPDF,pdfminer}] input_pdf_path

optional arguments:
  --output_bbox_img     Option to save text block images with low contrast.
  --output_dir OUTPUT_DIR
                        The directory for output images.
  --draw_bbox           Option to draw bounding boxes on low contrast text blocks.
  --output_pdf_path OUTPUT_PDF_PATH
                        The path for the output PDF file with drawn bounding boxes.
  --bbox_extractor {PyMyPDF,pdfminer}
                        Choice of bounding box extractor.

Default output folder: segmentation_model_output

Default output path with bounding box: bbox_low_contrast.pdf

• Extract text bounding boxes (PDFMiner)

python script/extract_text_bbox_PDFminer.py [--output_pdf_path OUTPUT_PDF_PATH] input_pdf_path

• Extract text bounding boxes (PyMuPDF)

python script/extract_text_bbox_PyMuPDF.py [--output_pdf_path OUTPUT_PDF_PATH] [--text_img] [--output_dir OUTPUT_DIR] input_pdf_path

Line Spacing

Analyze the line spacing using PDFminer

Criteria: Line height (line spacing) to at least 1.5 times the font size

WACG guideline: 1.4.12

Usage Instruction

python script/line_spacing.py input_pdf_path

WCAG Accessibility Issues Covered

• Images: non-text image (i.e. icon, header), image of text

• Text Presentation: line spacing, text-background contrast

• Language: language of page, language of parts

  PDF Language Detection

  Examines the PDF's metadata for a specified language property using Langdetect and PyMuPDF.

  Criteria: Assistive technology can determine the language of a page

  WACG guideline: 3.1.1, 3.1.2

  Basic usage:

  python script/language_detection.py input_pdf_path