image detection/extraction not work for CV paper https://arxiv.org/pdf/2111.06377

[genmini-ai]

logs

2025-09-17 04:19:48,800 - opencanvas.utils.plot_caption_extractor - INFO - Extracted plot from page 14, image 86 (383x128px)
2025-09-17 04:19:49,078 - opencanvas.utils.plot_caption_extractor - INFO - Extracted plot from page 14, image 87 (384x128px)
2025-09-17 04:19:49,323 - opencanvas.utils.plot_caption_extractor - INFO - Extracted plot from page 14, image 88 (384x128px)
2025-09-17 04:19:49,624 - opencanvas.utils.plot_caption_extractor - INFO - Extracted plot from page 14, image 89 (384x128px)
2025-09-17 04:19:49,843 - opencanvas.utils.plot_caption_extractor - INFO - Extracted plot from page 14, image 90 (384x128px)
2025-09-17 04:19:50,122 - opencanvas.utils.plot_caption_extractor - INFO - Extracted plot from page 14, image 91 (383x128px)
2025-09-17 04:19:50,376 - opencanvas.utils.plot_caption_extractor - INFO - Extracted plot from page 14, image 92 (383x128px)
2025-09-17 04:19:50,657 - opencanvas.utils.plot_caption_extractor - INFO - Extracted plot from page 14, image 93 (384x128px)
2025-09-17 04:19:50,898 - opencanvas.utils.plot_caption_extractor - INFO - Extracted plot from page 14, image 94 (384x128px)
2025-09-17 04:19:50,898 - opencanvas.utils.plot_caption_extractor - INFO - Extracted 250 plots from PDF

[genmini-ai]

Brainstorming with CC: Why Hundreds of Sub-Images Are Extracted

1. PDF Structure Issues

• Vector graphics broken into pieces: Some PDFs store charts/plots as hundreds of small vector elements
• Tiled images: Large images split into tiles for PDF optimization
• Background patterns: Watermarks or grid patterns counted as separate images
• Icon fonts: Each icon/symbol treated as an image

2. pdfplumber Detection Problems

• Over-segmentation: pdfplumber's page.images might detect:
  • Individual chart elements (bars, points, lines)
  • Legend items as separate images
  • Axis labels as images
  • Background gradients in pieces

3. PDF Generation Artifacts

• Print-to-PDF: Some tools create fragmented images
• PowerPoint/Excel exports: Often create many small image chunks
• Scanner artifacts: OCR software might segment incorrectly
• Compression artifacts: Some PDF compressors split images

4. Missing Filters/Validation

• No minimum size threshold (e.g., ignore images < 50x50 pixels)
• No aspect ratio filtering (e.g., skip very thin lines)
• No duplicate detection (same image at slightly different positions)
• No clustering of nearby fragments into single image
• No image type filtering (decorative vs actual plots)

5. Memory Multiplication Effect

• 100 sub-images × 300 DPI rendering = massive memory
• Each requires: to_image() → crop → PNG conversion → base64
• API calls for each fragment (if processing all)

6. Potential Solutions to Consider

• Pre-filtering: Size, aspect ratio, position clustering
• Image merging: Combine adjacent fragments before processing
• Sampling: Process only largest N images per page
• Page-level processing: Render entire page once, detect regions
• Different extraction library: PyMuPDF or Docling might handle better
• Heuristics: Detect "real" plots vs decorative elements

The key insights:

1. Root cause: PDFs fragmenting images into hundreds of sub-images
2. Memory explosion: Each fragment × 300 DPI × processing overhead
3. Missing filters: No size/relevance filtering before processing

The SIGKILL is likely from processing hundreds of fragments at high resolution without cleanup. Adding pre-filters and minimum size thresholds would prevent most of these
issues.