AI-Aware Image Compression with 11 Detectors

The short version: image compression for an AI pipeline must preserve downstream-detection features, not just shrink file size. AI-aware compression uses 11 detectors to mark critical regions and keeps faces and seals intact under high compression.

1. Why generic compression breaks things

Generic lossy compression cuts data by human visual redundancy, smoothing edges, textures, and fine structure. To a human it looks "no difference," but to OCR, face detection, and seal-recognition models the features are flattened and accuracy drops off a cliff. Screenshots, IDs, and seals in RAG documents often become unreadable after generic compression.

The mechanism is worth spelling out: detection models rely on gradients and local texture to locate targets, and quantization erases exactly those high-frequency details first. Human eyes are more sensitive to color than to edge position, so compression algorithms trade edges for color—the opposite of what models depend on. The result: a compressed sample batch looks fine to human reviewers, yet after bulk ingest the OCR error rate climbs, and only on specific content—seal strokes blur, ID digits bleed together. Teams usually suspect the model version first and circle back to compression parameters last.

2. Eleven detectors mark protected regions

AI-aware compression reuses the same 11 detectors as document desensitization: face, plate, text OSD, seal/signature, QR, screen, table, watermark logo, subtitle bar, ID segment, distant small object. Marked regions enter a protection list; compression lowers quantization strength there and preserves edges and texture, while everything else compresses normally.

Reusing one detector set pays off operationally: no separate model to maintain for compression, and the feature map is shared—a region that is both sensitive and critical (an official seal, say) is detected once and referenced twice. The difference is only in post-processing: desensitization blurs or boxes the region, feature protection quantizes it lightly. Opposite directions, identical detection layer.

3. Protection has a ceiling

More protection is not always better. The protected-region ratio is capped at 60%; beyond that it automatically falls back to mainline compression to avoid "protect everything" blowing up the file size. This is the engineering trade-off between feature preservation and size control.

Why fall back instead of forcing the protected regions to compress? Picture a full-page scanned table: the table area can cover seventy or eighty percent of the frame. Protecting all of it makes the compression-ratio target fail. Falling back to mainline compression processes the whole image under one policy, sacrificing some features to hold the size budget, with the SSIM / PSNR quality gate preventing over-compression distortion. The cap bounds the worst case—that is what makes it more reliable than clever auto-tuning.

4. Working with desensitization and parsing

AI-aware compression shares detectors with desensitization; the difference is intent—desensitization erases, feature protection preserves. Both intervene after parsing to ensure images entering RAG are both safe and machine-readable.

In the pipeline, the order reads: parse to get images, desensitize sensitive regions, compress with feature protection on what remains, then chunk and ingest. Compression uses the desensitized copy as source, so sensitive pixels never survive a second time inside compressed archives. One detection pass, two procedures, one quality gate—that is the full chain for images entering RAG.

Image handling is two sides of document preprocessing. Desensitization logic is in Why Desensitization Must Precede Embedding; parsing scans is in Scanned PDF OCR and Layout Analysis; full engine capability is in SDK and Compression.

Engineering Checklist: Balancing Feature Preservation and Size

• Cap the protected-area ratio: AI feature protection caps protected regions at 60% of the image, automatically falling back to the main compression pipeline beyond that. Without the cap, a scan full of sensitive regions becomes incompressible.
• Gate with SSIM / PSNR: even conservative settings must pass quality gates to prevent over-compression distorting the page layout — distorted layouts cause secondary errors in downstream OCR and layout analysis.
• Treat hard regions separately: subtitle bars, ID-number strips, and distant small objects are where both false positives and false misses concentrate; detector thresholds for these three differ from ordinary regions.
• Compress the desensitized copy: when compression runs integrated with preprocessing, its input prefers the desensitized copy — saving about 60% storage on average without introducing a second leak path.

All four points reduce to one idea: image desensitization and image compression are two gates on one pipeline, not two separate jobs. The full security chain is described in RAG Desensitization Flow.

FAQ

Q1: Why does generic compression hurt AI recognition?

Generic lossy compression cuts by human visual redundancy and smooths high-frequency edges first. Humans see no difference, but OCR, face, and seal models depend on exactly those features, so accuracy collapses—usually only on seals and ID content.

Q2: What marks regions for AI feature protection?

It reuses the same 11 desensitization detectors: face, plate, text OSD, seal/signature, QR, screen, table, watermark logo, subtitle bar, ID segment, distant small object. One detection pass, referenced by both procedures.

Q3: Why cap protected regions at 60%?

To avoid 'protect everything' blowing up file size. Beyond 60% it falls back to mainline compression with the SSIM / PSNR gate guarding against distortion, which bounds the worst case and balances features against size.

Q4: How does feature protection differ from desensitization?

They share detectors but opposite intent: desensitization erases sensitive pixels, feature protection preserves edges and texture in key regions so downstream models stay readable. Both run after parsing.