Skip to content

The aspect ratio of A4 paper is different from standard specifications. #95#96

Open
asylee02 wants to merge 2 commits intodatalab-to:devfrom
asylee02:master
Open

The aspect ratio of A4 paper is different from standard specifications. #95#96
asylee02 wants to merge 2 commits intodatalab-to:devfrom
asylee02:master

Conversation

@asylee02
Copy link
Copy Markdown

@asylee02 asylee02 commented Apr 30, 2026

Summary

When converting an A4 PDF, the rendered image has an aspect ratio of ~1.403
instead of the expected 1.4142, due to grid quantization in scale_to_fit.

Root Cause

Step 1 — input.py: load_pdf_images renders A4 correctly at 192 DPI:

  • 595.28 × 841.89 pt → 1587 × 2245 px → ratio 1.4147 ✓

Step 2 — model/util.py: scale_to_fit snaps to multiples of grid_size=28:

w_blocks = round(1587 / 28) = round(56.68) = 57  →  new_width  = 1596
h_blocks = round(2245 / 28) = round(80.18) = 80  →  new_height = 2240
Width rounds up (+9 px), height rounds down (−5 px) — both errors push
the aspect ratio in the same direction:

2240 / 1596 = 1.4035 ✗

Impact

The distortion is ~0.8%, which is negligible for OCR text accuracy, but causes
systematic error when mapping bounding box coordinates back to the original PDF
coordinate space.

Environment

- chandra-ocr v0.2.0
- IMAGE_DPI = 192, MIN_PDF_IMAGE_DIM = 1024, grid_size = 28

u-ashish and others added 2 commits April 22, 2026 16:05
@u-ashish
Dev
1742200
@asylee03
fix: preserve aspect ratio in scale_to_fit grid quantization
74c9bec 
  Independent rounding of w_blocks and h_blocks can push them in
  opposite directions (one up, one down), distorting the aspect ratio.
  For A4 at 192 DPI this produces 1.403 instead of 1.4142.

  Derive h_blocks from the already-rounded w_blocks so both axes
  are quantized consistently.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

3 participants

@asylee02 @u-ashish @asylee03