feat: OCR word_boxes fuer pixelgenaue Overlay-Positionierung

Backend: _ocr_cell_crop speichert jetzt word_boxes mit exakten
Tesseract/RapidOCR Wort-Koordinaten (left, top, width, height)
im Cell-Ergebnis. Absolute Bildkoordinaten, bereits zurueckgemappt.

Frontend: Slide-Hook nutzt word_boxes direkt wenn vorhanden —
jedes Wort wird exakt an seiner OCR-Position platziert. Kein
Pixel-Scanning noetig. Fallback auf alten Slide wenn keine Boxes.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

admin-lehrer/app/(admin)/ai/ocr-overlay/types.ts

@@ -16,6 +16,7 @@ export type {
   RowItem,
   GridResult,
   GridCell,
+  OcrWordBox,
   WordBbox,
   ColumnMeta,
 } from '../ocr-pipeline/types'

admin-lehrer/app/(admin)/ai/ocr-pipeline/types.ts

@@ -220,6 +220,15 @@ export interface WordBbox {
   h: number
 }
 
+export interface OcrWordBox {
+  text: string
+  left: number    // absolute image x in px
+  top: number     // absolute image y in px
+  width: number   // px
+  height: number  // px
+  conf: number
+}
+
 export interface GridCell {
   cell_id: string          // "R03_C1"
   row_index: number
@@ -232,6 +241,7 @@ export interface GridCell {
   ocr_engine?: string
   is_bold?: boolean
   status?: 'pending' | 'confirmed' | 'edited' | 'skipped'
+  word_boxes?: OcrWordBox[]  // per-word bounding boxes from OCR engine
 }
 
 export interface ColumnMeta {

admin-lehrer/components/ocr-overlay/useSlideWordPositions.ts

@@ -9,16 +9,16 @@ export interface WordPosition {
 }
 
 /**
- * "Slide from left" positioning algorithm.
+ * "Slide from left" positioning using OCR word bounding boxes.
  *
- * Takes ALL recognised words per cell and slides them left-to-right across
+ * If the backend provides `word_boxes` (exact per-word coordinates from
- * the row's dark-pixel projection until each word "locks" onto its ink.
+ * Tesseract/RapidOCR), we place each word directly at its OCR position.
+ * This gives pixel-accurate overlay without any heuristic pixel scanning.
  *
- * Font size: fontRatio = 1.0 for all tokens (matches fallback rendering).
+ * Fallback: if no word_boxes, slide tokens across dark-pixel projection
- * Token widths: derived from canvas measureText scaled to the rendered font
+ * (original slide algorithm).
- * size (medianCh * 0.7), ensuring visually correct proportions.
  *
- * Guarantees: no words dropped, no complex matching rules needed.
+ * Font size: fontRatio = 1.0 for all (matches fallback rendering).
  */
 export function useSlideWordPositions(
   imageUrl: string,
@@ -37,6 +37,37 @@ export function useSlideWordPositions(
       const imgW = img.naturalWidth
       const imgH = img.naturalHeight
 
+      // Check if we can use word_boxes (fast path — no canvas needed)
+      const hasWordBoxes = cells.some(c => c.word_boxes && c.word_boxes.length > 0)
+
+      if (hasWordBoxes) {
+        // --- FAST PATH: use OCR word bounding boxes directly ---
+        const positions = new Map<string, WordPosition[]>()
+
+        for (const cell of cells) {
+          if (!cell.bbox_pct || !cell.text) continue
+          const boxes = cell.word_boxes
+          if (!boxes || boxes.length === 0) continue
+
+          const wordPos: WordPosition[] = boxes
+            .filter(wb => wb.text.trim())
+            .map(wb => ({
+              xPct: (wb.left / imgW) * 100,
+              wPct: (wb.width / imgW) * 100,
+              text: wb.text,
+              fontRatio: 1.0,
+            }))
+
+          if (wordPos.length > 0) {
+            positions.set(cell.cell_id, wordPos)
+          }
+        }
+
+        setResult(positions)
+        return
+      }
+
+      // --- SLOW PATH: pixel-projection slide (fallback if no word_boxes) ---
       const canvas = document.createElement('canvas')
       canvas.width = imgW
       canvas.height = imgH
@@ -56,7 +87,6 @@ export function useSlideWordPositions(
       const fontFam = "'Liberation Sans', Arial, sans-serif"
       ctx.font = `${refFontSize}px ${fontFam}`
 
-      // --- Global font scale from median cell height ---
       const cellHeights = cells
         .filter(c => c.bbox_pct && c.bbox_pct.h > 0)
         .map(c => Math.round(c.bbox_pct.h / 100 * imgH))
@@ -65,7 +95,6 @@ export function useSlideWordPositions(
         ? cellHeights[Math.floor(cellHeights.length / 2)]
         : 30
 
-      // measureScale maps measureText pixels → image pixels at rendered font
       const renderedFontImgPx = medianCh * 0.7
       const measureScale = renderedFontImgPx / refFontSize
       const spaceWidthPx = Math.max(2, Math.round(ctx.measureText(' ').width * measureScale))
@@ -91,7 +120,6 @@ export function useSlideWordPositions(
         if (cy < 0) cy = 0
         if (cx + cw > imgW || cy + ch > imgH) continue
 
-        // --- Dark-pixel projection ---
         const imageData = ctx.getImageData(cx, cy, cw, ch)
         const proj = new Float32Array(cw)
         for (let y = 0; y < ch; y++) {
@@ -111,23 +139,18 @@ export function useSlideWordPositions(
           ink.reverse()
         }
 
-        // --- Split into individual tokens ---
         const tokens = cell.text.split(/\s+/).filter(Boolean)
         if (tokens.length === 0) continue
 
-        // Token widths in image pixels (at rendered font size)
         const tokenWidthsPx = tokens.map(t =>
           Math.max(4, Math.round(ctx.measureText(t).width * measureScale))
         )
 
-        // --- Slide each token left-to-right ---
         const wordPos: WordPosition[] = []
         let cursor = 0
 
         for (let ti = 0; ti < tokens.length; ti++) {
           const tokenW = tokenWidthsPx[ti]
-
-          // Find first x from cursor where ≥15% of span has ink
           const coverageNeeded = Math.max(1, Math.round(tokenW * 0.15))
           let bestX = cursor
 
@@ -143,14 +166,12 @@ export function useSlideWordPositions(
               bestX = x
               break
             }
-            // Safety: don't scan more than 30% of cell width past cursor
             if (x > cursor + cw * 0.3 && ti > 0) {
               bestX = cursor
               break
             }
           }
 
-          // Clamp to cell bounds
           if (bestX + tokenW > cw) {
             bestX = Math.max(0, cw - tokenW)
           }
@@ -162,7 +183,6 @@ export function useSlideWordPositions(
             fontRatio: 1.0,
           })
 
-          // Advance cursor: past this token + space
           cursor = bestX + tokenW + spaceWidthPx
         }
 

klausur-service/backend/cv_cell_grid.py

@@ -221,6 +221,20 @@ def _ocr_cell_crop(
                     sum(w['conf'] for w in psm7_words) / len(psm7_words), 1
                 )
                 used_engine = 'cell_crop_v2_psm7'
+                # Remap PSM7 word positions back to original image coords
+                if up_w != cw or up_h != ch:
+                    sx = cw / max(up_w, 1)
+                    sy = ch / max(up_h, 1)
+                    for w in psm7_words:
+                        w['left'] = int(w['left'] * sx) + cx
+                        w['top'] = int(w['top'] * sy) + cy
+                        w['width'] = int(w['width'] * sx)
+                        w['height'] = int(w['height'] * sy)
+                else:
+                    for w in psm7_words:
+                        w['left'] += cx
+                        w['top'] += cy
+                words = psm7_words
 
     # --- Noise filter ---
     if text.strip():
@@ -235,6 +249,23 @@ def _ocr_cell_crop(
     result['text'] = text
     result['confidence'] = avg_conf
     result['ocr_engine'] = used_engine
+
+    # Store individual word bounding boxes (absolute image coordinates)
+    # for pixel-accurate overlay positioning in the frontend.
+    if words and text.strip():
+        result['word_boxes'] = [
+            {
+                'text': w.get('text', ''),
+                'left': w['left'],
+                'top': w['top'],
+                'width': w['width'],
+                'height': w['height'],
+                'conf': w.get('conf', 0),
+            }
+            for w in words
+            if w.get('text', '').strip()
+        ]
+
     return result