63ffa8f0e3
Feature implementation completed and tested successfully. - PP-StructureV3 orientation detection enabled - Page dimensions correctly swapped for 90°/270° rotations - Output PDF now displays landscape content correctly 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
53 lines
2.7 KiB
Markdown
53 lines
2.7 KiB
Markdown
# Enable Document Orientation Detection
|
|
|
|
## Summary
|
|
Enable PP-StructureV3's document orientation classification feature to correctly handle PDF scans where the content orientation differs from the PDF page metadata.
|
|
|
|
## Problem Statement
|
|
Currently, when a portrait-oriented PDF contains landscape-scanned content (or vice versa), the OCR system produces incorrect results because:
|
|
|
|
1. **pdf2image** extracts images based on PDF metadata (e.g., `Page size: 1242 x 1755`, `Page rot: 0`)
|
|
2. **PP-StructureV3** has `use_doc_orientation_classify=False` (disabled)
|
|
3. The OCR attempts to read sideways text, resulting in poor recognition
|
|
4. The output PDF has wrong page dimensions
|
|
|
|
### Example Scenario
|
|
- Input: Portrait PDF (1242 x 1755) containing landscape-scanned delivery form
|
|
- Current output: Portrait PDF with unreadable/incorrect text
|
|
- Expected output: Landscape PDF (1755 x 1242) with correctly oriented text
|
|
|
|
## Proposed Solution
|
|
Enable document orientation detection in PP-StructureV3 and adjust page dimensions based on the detected rotation:
|
|
|
|
1. **Enable orientation detection**: Set `use_doc_orientation_classify=True` in config
|
|
2. **Capture rotation info**: Extract the detected rotation angle (0°/90°/180°/270°) from PP-StructureV3 results
|
|
3. **Adjust dimensions**: When 90° or 270° rotation is detected, swap width and height for the output PDF
|
|
4. **Use OCR coordinates directly**: PP-StructureV3 returns coordinates based on the rotated image, so no coordinate transformation is needed
|
|
|
|
## PP-StructureV3 Orientation Detection Details
|
|
According to PaddleOCR documentation:
|
|
- **Stage 1 preprocessing**: `use_doc_orientation_classify` detects and rotates the entire page
|
|
- **Output format**: `doc_preprocessor_res` contains:
|
|
- `class_ids`: [0-3] corresponding to [0°, 90°, 180°, 270°]
|
|
- `label_names`: ["0", "90", "180", "270"]
|
|
- `scores`: confidence scores
|
|
- **Model accuracy**: PP-LCNet_x1_0_doc_ori achieves 99.06% top-1 accuracy
|
|
|
|
## Scope
|
|
- Backend only (no frontend changes required)
|
|
- Affects OCR track processing
|
|
- Does not affect Direct or Hybrid track
|
|
|
|
## Risks and Mitigations
|
|
| Risk | Mitigation |
|
|
|------|------------|
|
|
| Model might incorrectly classify mixed-orientation pages | 99.06% accuracy is acceptable; `use_textline_orientation` (already enabled) handles per-line correction |
|
|
| Coordinate mismatch in edge cases | Thorough testing with portrait, landscape, and mixed documents |
|
|
| Performance overhead | Orientation classification adds ~100ms per page (negligible vs total OCR time) |
|
|
|
|
## Success Criteria
|
|
1. Portrait PDF with landscape content produces landscape output PDF
|
|
2. Landscape PDF with portrait content produces portrait output PDF
|
|
3. Normal orientation documents continue to work correctly
|
|
4. Text recognition accuracy improves for rotated documents
|