ML Pipeline & Inpainting Engine

XianScan executes a modular, multi-threaded neural pipeline implemented directly in native Rust:

[Raw Webtoon Strip / Comic Page]
   │
   ▼
1. Webtoon Gutter Reslicing (src/ml/reslice.rs)
   │ ── Detects non-text whitespace valleys and avoids slicing panels/bubbles
   ▼
2. Bubble & Text Line Detection (src/ml/detect/)
   │ ── Koharu RF-DETR Seg 2XL & DBNet extract oriented polygon contours
   ▼
3. Multi-Script OCR & Reading Flow Fusion (src/ml/ocr/ & src/pipeline/fusion.rs)
   │ ── RapidOCR (10 languages) with cultural flow sorting (RTL Manga / LTR Webtoons)
   ▼
4. Context-Aware LLM Translation & Glossaries (web/src/lib/server/translate/)
   │ ── 5-page sliding dialogue tracker + Aho-Corasick multilingual domain dictionaries
   ▼
5. Neural Background Inpainting (src/ml/inpaint/)
   │ ── 1:1 localized patch cropping with LaMa Fast Fourier Convolutions (FFC)
   ▼
6. Automated Typography & Typesetting (web/src/lib/server/typeset/)
   │ ── Google Skia binary-search font sizing, outline strokes, and bubble tilt
   ▼
[Cleaned & Typeset Comic Canvas]

• Koharu RF-DETR Seg 2XL (src/ml/detect/rfdetr.rs): High-resolution transformer segmentation model that outputs exact polygon boundaries for oval speech bubbles, jagged scream bubbles, and rectangular narrative caption boxes.
• DBNet Polygon Extraction (src/ml/detect/dbnet.rs): Differentiable Binarization text detector capturing fine-grained character bounding coordinates and orientation angles.
• Spatial Clustering & Deduplication (src/pipeline/region_builder/):
• clustering.rs: Aggregates overlapping or adjacent text lines into coherent dialogue blocks.
• expansion.rs: Expands text bounding boxes with configurable safety padding to ensure full glyph coverage.
• dedup.rs: Resolves conflicting bounding boxes using Intersection-over-Union (IoU) heuristics.

• RapidOCR PP-OCRv6 Engine (src/ml/ocr/engine.rs): Embedded multi-lingual OCR pipeline supporting 10 languages:
• Chinese (Simplified & Traditional), Japanese (Kanji, Hiragana, Katakana), Korean (Hangul), Thai, Indonesian, English, Spanish, French, Russian.
• Cultural Reading Flow Sorter (src/ml/detect/grouping.rs & src/pipeline/fusion.rs):
• Japanese Manga: Right-to-Left (RTL) reading flow. Iterates vertical text columns from right to left within horizontal narrative bands.
• Korean Manhwa & Chinese Manhua: Top-to-Bottom vertical flow with left-to-right reading hierarchy.

• Large Mask Inpainting (LaMa) Architecture (src/ml/inpaint/lama.rs): Utilizes Fast Fourier Convolutions (FFC) with global receptive fields to reconstruct complex background textures, screentone patterns, and gradients.
• 1:1 Localized Patch Cropping Engine (src/ml/inpaint/patch.rs):
• Traditional inpainting tools downscale entire 4K comic canvases to 512x512, causing severe blurriness across the whole page.
• XianScan crops tight local patches around detected dialogue bubbles, inpaints each patch at native resolution, and composites them back seamlessly onto the unedited source image.
• Solid Background Fast Path (is_solid_background_patch):
• Automatically identifies solid / flat white speech bubbles and replaces them instantly with exact color fills, bypassing GPU neural execution for sub-millisecond performance.

• Google Skia Canvas Engine (@napi-rs/canvas): Native high-performance 2D graphics rendering embedded directly in the server.
• Binary Search Font Sizing (web/src/lib/server/typeset/layout.ts): Dynamically computes the maximum legible font size that fits comfortably within the speech bubble's polygon mask.
• Multi-Line Text Wrapping & Hyphenation: Balances line lengths to prevent orphan words and maintain aesthetic dialogue shapes.
• Outline Strokes & Bubble Tilt: Applies configurable contrast outline strokes and rotates text lines to match the orientation angle of tilted dialogue bubbles.