feat: Sprint 2 — TrOCR ONNX, PP-DocLayout, Model Management

D2: TrOCR ONNX export script (printed + handwritten, int8 quantization)
D3: PP-DocLayout ONNX export script (download or Docker-based conversion)
B3: Model Management admin page (PyTorch vs ONNX status, benchmarks, config)
A4: TrOCR ONNX service with runtime routing (auto/pytorch/onnx via TROCR_BACKEND)
A5: PP-DocLayout ONNX detection with OpenCV fallback (via GRAPHIC_DETECT_BACKEND)
B4: Structure Detection UI toggle (OpenCV vs PP-DocLayout) with class color coding
C3: TrOCR-ONNX.md documentation
C4: OCR-Pipeline.md ONNX section added
C5: mkdocs.yml nav updated, optimum added to requirements.txt

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
Benjamin Admin
2026-03-23 09:53:02 +01:00
parent c695b659fb
commit be7f5f1872
16 changed files with 3616 additions and 60 deletions

View File

@@ -233,6 +233,15 @@ export interface ExcludeRegion {
label?: string
}
export interface DocLayoutRegion {
x: number
y: number
w: number
h: number
class_name: string
confidence: number
}
export interface StructureResult {
image_width: number
image_height: number
@@ -246,6 +255,9 @@ export interface StructureResult {
word_count: number
border_ghosts_removed?: number
duration_seconds: number
/** PP-DocLayout regions (only present when method=ppdoclayout) */
layout_regions?: DocLayoutRegion[]
detection_method?: 'opencv' | 'ppdoclayout'
}
export interface StructureBox {