feat(klausur): Handschrift entfernen + Klausur-HTR implementiert

Feature 1: Handschrift entfernen via OCR-Pipeline Session
- services/handwriting_detection.py: _detect_pencil() + target_ink Parameter
  ("all" | "colored" | "pencil") für gezielte Tinten-Erkennung
- ocr_pipeline_session_store.py: clean_png + handwriting_removal_meta Spalten
  (idempotentes ALTER TABLE in init_ocr_pipeline_tables)
- ocr_pipeline_api.py: POST /sessions/{id}/remove-handwriting Endpoint
  + "clean" zu valid_types für Image-Serving hinzugefügt

Feature 2: Klausur-HTR (Hochwertige Handschriftenerkennung)
- handwriting_htr_api.py: Neuer Router /api/v1/htr/recognize + /recognize-session
  Primary: qwen2.5vl:32b via Ollama, Fallback: trocr-large-handwritten
- services/trocr_service.py: size Parameter (base | large) für get_trocr_model()
  + run_trocr_ocr() - unterstützt jetzt trocr-large-handwritten
- main.py: HTR Router registriert

Config:
- docker-compose.yml: OLLAMA_HTR_MODEL, HTR_FALLBACK_MODEL
- .env.example: HTR Env-Vars dokumentiert

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

.env.example

@@ -30,6 +30,23 @@ OLLAMA_VISION_MODEL=llama3.2-vision
 OLLAMA_CORRECTION_MODEL=llama3.2
 OLLAMA_TIMEOUT=120
 
+# OCR-Pipeline: LLM-Review (Schritt 6)
+# Kleine Modelle reichen fuer Zeichen-Korrekturen (0->O, 1->l, 5->S)
+# Optionen: qwen3:0.6b, qwen3:1.7b, gemma3:1b, qwen3:30b-a3b
+OLLAMA_REVIEW_MODEL=qwen3:0.6b
+# Eintraege pro Ollama-Call. Groesser = weniger HTTP-Overhead.
+OLLAMA_REVIEW_BATCH_SIZE=20
+
+# OCR-Pipeline: Engine fuer Schritt 5 (Worterkennung)
+# Optionen: auto (bevorzugt RapidOCR), rapid, tesseract,
+#           trocr-printed, trocr-handwritten, lighton
+OCR_ENGINE=auto
+
+# Klausur-HTR: Primaerem Modell fuer Handschriftenerkennung (qwen2.5vl bereits auf Mac Mini)
+OLLAMA_HTR_MODEL=qwen2.5vl:32b
+# HTR Fallback: genutzt wenn Ollama nicht erreichbar (auto-download ~340 MB)
+HTR_FALLBACK_MODEL=trocr-large
+
 # Anthropic (optional)
 ANTHROPIC_API_KEY=
 

docker-compose.yml

@@ -15,6 +15,7 @@ volumes:
   eh_uploads:
   ocr_labeling:
   paddle_models:
+  lighton_models:
   paddleocr_models:
   transcription_models:
   transcription_temp:
@@ -209,6 +210,7 @@ services:
       - eh_uploads:/app/eh-uploads
       - ocr_labeling:/app/ocr-labeling
       - paddle_models:/root/.paddlex
+      - lighton_models:/root/.cache/huggingface
     environment:
       JWT_SECRET: ${JWT_SECRET:-your-super-secret-jwt-key-change-in-production}
       BACKEND_URL: http://backend-lehrer:8001
@@ -231,6 +233,11 @@ services:
       OLLAMA_DEFAULT_MODEL: ${OLLAMA_DEFAULT_MODEL:-llama3.2}
       OLLAMA_VISION_MODEL: ${OLLAMA_VISION_MODEL:-llama3.2-vision}
       OLLAMA_CORRECTION_MODEL: ${OLLAMA_CORRECTION_MODEL:-llama3.2}
+      OLLAMA_REVIEW_MODEL: ${OLLAMA_REVIEW_MODEL:-qwen3:0.6b}
+      OLLAMA_REVIEW_BATCH_SIZE: ${OLLAMA_REVIEW_BATCH_SIZE:-20}
+      OCR_ENGINE: ${OCR_ENGINE:-auto}
+      OLLAMA_HTR_MODEL: ${OLLAMA_HTR_MODEL:-qwen2.5vl:32b}
+      HTR_FALLBACK_MODEL: ${HTR_FALLBACK_MODEL:-trocr-large}
       RAG_SERVICE_URL: http://bp-core-rag-service:8097
     extra_hosts:
       - "host.docker.internal:host-gateway"

klausur-service/backend/handwriting_htr_api.py

@@ -0,0 +1,276 @@
+"""
+Handwriting HTR API - Hochwertige Handschriftenerkennung (HTR) fuer Klausurkorrekturen.
+
+Endpoints:
+- POST /api/v1/htr/recognize          - Bild hochladen → handgeschriebener Text
+- POST /api/v1/htr/recognize-session  - OCR-Pipeline Session als Quelle nutzen
+
+Modell-Strategie:
+  1. qwen2.5vl:32b via Ollama (primaer, hoechste Qualitaet als VLM)
+  2. microsoft/trocr-large-handwritten (Fallback, offline, kein Ollama)
+
+DATENSCHUTZ: Alle Verarbeitung erfolgt lokal auf dem Mac Mini.
+"""
+
+import io
+import os
+import logging
+import time
+import base64
+from typing import Optional
+
+import cv2
+import numpy as np
+from fastapi import APIRouter, HTTPException, Query, UploadFile, File
+from pydantic import BaseModel
+
+logger = logging.getLogger(__name__)
+
+router = APIRouter(prefix="/api/v1/htr", tags=["HTR"])
+
+OLLAMA_BASE_URL = os.getenv("OLLAMA_BASE_URL", "http://host.docker.internal:11434")
+OLLAMA_HTR_MODEL = os.getenv("OLLAMA_HTR_MODEL", "qwen2.5vl:32b")
+HTR_FALLBACK_MODEL = os.getenv("HTR_FALLBACK_MODEL", "trocr-large")
+
+
+# ---------------------------------------------------------------------------
+# Pydantic Models
+# ---------------------------------------------------------------------------
+
+class HTRSessionRequest(BaseModel):
+    session_id: str
+    model: str = "auto"       # "auto" | "qwen2.5vl" | "trocr-large"
+    use_clean: bool = True    # Prefer clean_png (after handwriting removal)
+
+
+# ---------------------------------------------------------------------------
+# Preprocessing
+# ---------------------------------------------------------------------------
+
+def _preprocess_for_htr(img_bgr: np.ndarray) -> np.ndarray:
+    """
+    CLAHE contrast enhancement + upscale to improve HTR accuracy.
+    Returns grayscale enhanced image.
+    """
+    gray = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2GRAY)
+    clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8, 8))
+    enhanced = clahe.apply(gray)
+
+    # Upscale if image is too small
+    h, w = enhanced.shape
+    if min(h, w) < 800:
+        scale = 800 / min(h, w)
+        enhanced = cv2.resize(
+            enhanced, None, fx=scale, fy=scale,
+            interpolation=cv2.INTER_CUBIC
+        )
+
+    return enhanced
+
+
+def _bgr_to_png_bytes(img_bgr: np.ndarray) -> bytes:
+    """Convert BGR ndarray to PNG bytes."""
+    success, buf = cv2.imencode(".png", img_bgr)
+    if not success:
+        raise RuntimeError("Failed to encode image to PNG")
+    return buf.tobytes()
+
+
+def _preprocess_image_bytes(image_bytes: bytes) -> bytes:
+    """Load image, apply HTR preprocessing, return PNG bytes."""
+    arr = np.frombuffer(image_bytes, dtype=np.uint8)
+    img_bgr = cv2.imdecode(arr, cv2.IMREAD_COLOR)
+    if img_bgr is None:
+        raise ValueError("Could not decode image")
+
+    enhanced = _preprocess_for_htr(img_bgr)
+    # Convert grayscale back to BGR for encoding
+    enhanced_bgr = cv2.cvtColor(enhanced, cv2.COLOR_GRAY2BGR)
+    return _bgr_to_png_bytes(enhanced_bgr)
+
+
+# ---------------------------------------------------------------------------
+# Backend: Ollama qwen2.5vl
+# ---------------------------------------------------------------------------
+
+async def _recognize_with_qwen_vl(image_bytes: bytes, language: str) -> Optional[str]:
+    """
+    Send image to Ollama qwen2.5vl:32b for HTR.
+    Returns extracted text or None on error.
+    """
+    import httpx
+
+    lang_hint = {
+        "de": "Deutsch",
+        "en": "Englisch",
+        "de+en": "Deutsch und Englisch",
+    }.get(language, "Deutsch")
+
+    prompt = (
+        f"Du bist ein OCR-Experte fuer handgeschriebenen Text auf {lang_hint}. "
+        "Lies den Text im Bild exakt ab — korrigiere KEINE Rechtschreibfehler. "
+        "Antworte NUR mit dem erkannten Text, ohne Erklaerungen."
+    )
+
+    img_b64 = base64.b64encode(image_bytes).decode("utf-8")
+
+    payload = {
+        "model": OLLAMA_HTR_MODEL,
+        "prompt": prompt,
+        "images": [img_b64],
+        "stream": False,
+    }
+
+    try:
+        async with httpx.AsyncClient(timeout=120.0) as client:
+            resp = await client.post(f"{OLLAMA_BASE_URL}/api/generate", json=payload)
+            resp.raise_for_status()
+            data = resp.json()
+            return data.get("response", "").strip()
+    except Exception as e:
+        logger.warning(f"Ollama qwen2.5vl HTR failed: {e}")
+        return None
+
+
+# ---------------------------------------------------------------------------
+# Backend: TrOCR-large fallback
+# ---------------------------------------------------------------------------
+
+async def _recognize_with_trocr_large(image_bytes: bytes) -> Optional[str]:
+    """
+    Use microsoft/trocr-large-handwritten via trocr_service.py.
+    Returns extracted text or None on error.
+    """
+    try:
+        from services.trocr_service import run_trocr_ocr, _check_trocr_available
+        if not _check_trocr_available():
+            logger.warning("TrOCR not available for HTR fallback")
+            return None
+
+        text, confidence = await run_trocr_ocr(image_bytes, handwritten=True, size="large")
+        return text.strip() if text else None
+    except Exception as e:
+        logger.warning(f"TrOCR-large HTR failed: {e}")
+        return None
+
+
+# ---------------------------------------------------------------------------
+# Core recognition logic
+# ---------------------------------------------------------------------------
+
+async def _do_recognize(
+    image_bytes: bytes,
+    model: str = "auto",
+    preprocess: bool = True,
+    language: str = "de",
+) -> dict:
+    """
+    Core HTR logic: preprocess → try Ollama → fallback to TrOCR-large.
+    Returns dict with text, model_used, processing_time_ms.
+    """
+    t0 = time.monotonic()
+
+    if preprocess:
+        try:
+            image_bytes = _preprocess_image_bytes(image_bytes)
+        except Exception as e:
+            logger.warning(f"HTR preprocessing failed, using raw image: {e}")
+
+    text: Optional[str] = None
+    model_used: str = "none"
+
+    use_qwen = model in ("auto", "qwen2.5vl")
+    use_trocr = model in ("auto", "trocr-large") or (use_qwen and text is None)
+
+    if use_qwen:
+        text = await _recognize_with_qwen_vl(image_bytes, language)
+        if text is not None:
+            model_used = f"qwen2.5vl ({OLLAMA_HTR_MODEL})"
+
+    if text is None and (use_trocr or model == "trocr-large"):
+        text = await _recognize_with_trocr_large(image_bytes)
+        if text is not None:
+            model_used = "trocr-large-handwritten"
+
+    if text is None:
+        text = ""
+        model_used = "none (all backends failed)"
+
+    elapsed_ms = int((time.monotonic() - t0) * 1000)
+
+    return {
+        "text": text,
+        "model_used": model_used,
+        "processing_time_ms": elapsed_ms,
+        "language": language,
+        "preprocessed": preprocess,
+    }
+
+
+# ---------------------------------------------------------------------------
+# Endpoints
+# ---------------------------------------------------------------------------
+
+@router.post("/recognize")
+async def recognize_handwriting(
+    file: UploadFile = File(...),
+    model: str = Query("auto", description="auto | qwen2.5vl | trocr-large"),
+    preprocess: bool = Query(True, description="Apply CLAHE + upscale before recognition"),
+    language: str = Query("de", description="de | en | de+en"),
+):
+    """
+    Upload an image and get back the handwritten text as plain text.
+
+    Tries qwen2.5vl:32b via Ollama first, falls back to TrOCR-large-handwritten.
+    """
+    if model not in ("auto", "qwen2.5vl", "trocr-large"):
+        raise HTTPException(status_code=400, detail="model must be one of: auto, qwen2.5vl, trocr-large")
+    if language not in ("de", "en", "de+en"):
+        raise HTTPException(status_code=400, detail="language must be one of: de, en, de+en")
+
+    image_bytes = await file.read()
+    if not image_bytes:
+        raise HTTPException(status_code=400, detail="Empty file")
+
+    return await _do_recognize(image_bytes, model=model, preprocess=preprocess, language=language)
+
+
+@router.post("/recognize-session")
+async def recognize_from_session(req: HTRSessionRequest):
+    """
+    Use an OCR-Pipeline session as image source for HTR.
+
+    Set use_clean=true to prefer the clean image (after handwriting removal step).
+    This is useful when you want to do HTR on isolated handwriting regions.
+    """
+    from ocr_pipeline_session_store import get_session_db, get_session_image
+
+    session = await get_session_db(req.session_id)
+    if not session:
+        raise HTTPException(status_code=404, detail=f"Session {req.session_id} not found")
+
+    # Choose source image
+    image_bytes: Optional[bytes] = None
+    source_used: str = ""
+
+    if req.use_clean:
+        image_bytes = await get_session_image(req.session_id, "clean")
+        if image_bytes:
+            source_used = "clean"
+
+    if not image_bytes:
+        image_bytes = await get_session_image(req.session_id, "deskewed")
+        if image_bytes:
+            source_used = "deskewed"
+
+    if not image_bytes:
+        image_bytes = await get_session_image(req.session_id, "original")
+        source_used = "original"
+
+    if not image_bytes:
+        raise HTTPException(status_code=404, detail="No image available in session")
+
+    result = await _do_recognize(image_bytes, model=req.model)
+    result["session_id"] = req.session_id
+    result["source_image"] = source_used
+    return result

klausur-service/backend/main.py

@@ -44,6 +44,10 @@ except ImportError:
 from vocab_worksheet_api import router as vocab_router, set_db_pool as set_vocab_db_pool, _init_vocab_table, _load_all_sessions, DATABASE_URL as VOCAB_DATABASE_URL
 from ocr_pipeline_api import router as ocr_pipeline_router
 from ocr_pipeline_session_store import init_ocr_pipeline_tables
+try:
+    from handwriting_htr_api import router as htr_router
+except ImportError:
+    htr_router = None
 try:
     from dsfa_rag_api import router as dsfa_rag_router, set_db_pool as set_dsfa_db_pool
     from dsfa_corpus_ingestion import DSFAQdrantService, DATABASE_URL as DSFA_DATABASE_URL
@@ -113,6 +117,19 @@ async def lifespan(app: FastAPI):
     # Ensure EH upload directory exists
     os.makedirs(EH_UPLOAD_DIR, exist_ok=True)
 
+    # Preload LightOnOCR model if OCR_ENGINE=lighton (avoids cold-start on first request)
+    ocr_engine_env = os.getenv("OCR_ENGINE", "auto")
+    if ocr_engine_env == "lighton":
+        try:
+            import asyncio
+            from services.lighton_ocr_service import get_lighton_model
+            loop = asyncio.get_event_loop()
+            print("Preloading LightOnOCR-2-1B at startup (OCR_ENGINE=lighton)...")
+            await loop.run_in_executor(None, get_lighton_model)
+            print("LightOnOCR-2-1B preloaded")
+        except Exception as e:
+            print(f"Warning: LightOnOCR preload failed: {e}")
+
     yield
 
     print("Klausur-Service shutting down...")
@@ -160,6 +177,8 @@ if trocr_router:
     app.include_router(trocr_router)      # TrOCR Handwriting OCR
 app.include_router(vocab_router)      # Vocabulary Worksheet Generator
 app.include_router(ocr_pipeline_router)  # OCR Pipeline (step-by-step)
+if htr_router:
+    app.include_router(htr_router)            # Handwriting HTR (Klausur)
 if dsfa_rag_router:
     app.include_router(dsfa_rag_router)   # DSFA RAG Corpus Search
 

klausur-service/backend/ocr_pipeline_api.py

@@ -168,6 +168,13 @@ class RowGroundTruthRequest(BaseModel):
     notes: Optional[str] = None
 
 
+class RemoveHandwritingRequest(BaseModel):
+    method: str = "auto"       # "auto" | "telea" | "ns"
+    target_ink: str = "all"    # "all" | "colored" | "pencil"
+    dilation: int = 2          # mask dilation iterations (0-5)
+    use_source: str = "auto"   # "original" | "deskewed" | "auto"
+
+
 # ---------------------------------------------------------------------------
 # Session Management Endpoints
 # ---------------------------------------------------------------------------
@@ -309,7 +316,7 @@ async def delete_session(session_id: str):
 @router.get("/sessions/{session_id}/image/{image_type}")
 async def get_image(session_id: str, image_type: str):
     """Serve session images: original, deskewed, dewarped, binarized, columns-overlay, or rows-overlay."""
-    valid_types = {"original", "deskewed", "dewarped", "binarized", "columns-overlay", "rows-overlay", "words-overlay"}
+    valid_types = {"original", "deskewed", "dewarped", "binarized", "columns-overlay", "rows-overlay", "words-overlay", "clean"}
     if image_type not in valid_types:
         raise HTTPException(status_code=400, detail=f"Unknown image type: {image_type}")
 
@@ -1906,3 +1913,90 @@ async def _get_words_overlay(session_id: str) -> Response:
         raise HTTPException(status_code=500, detail="Failed to encode overlay image")
 
     return Response(content=result_png.tobytes(), media_type="image/png")
+
+
+# ---------------------------------------------------------------------------
+# Handwriting Removal Endpoint
+# ---------------------------------------------------------------------------
+
+@router.post("/sessions/{session_id}/remove-handwriting")
+async def remove_handwriting_endpoint(session_id: str, req: RemoveHandwritingRequest):
+    """
+    Remove handwriting from a session image using inpainting.
+
+    Steps:
+    1. Load source image (auto → deskewed if available, else original)
+    2. Detect handwriting mask (filtered by target_ink)
+    3. Dilate mask to cover stroke edges
+    4. Inpaint the image
+    5. Store result as clean_png in the session
+
+    Returns metadata including the URL to fetch the clean image.
+    """
+    import time as _time
+    t0 = _time.monotonic()
+
+    from services.handwriting_detection import detect_handwriting
+    from services.inpainting_service import inpaint_image, dilate_mask as _dilate_mask, InpaintingMethod, image_to_png
+
+    session = await get_session_db(session_id)
+    if not session:
+        raise HTTPException(status_code=404, detail=f"Session {session_id} not found")
+
+    # 1. Determine source image
+    source = req.use_source
+    if source == "auto":
+        deskewed = await get_session_image(session_id, "deskewed")
+        source = "deskewed" if deskewed else "original"
+
+    image_bytes = await get_session_image(session_id, source)
+    if not image_bytes:
+        raise HTTPException(status_code=404, detail=f"Source image '{source}' not available")
+
+    # 2. Detect handwriting mask
+    detection = detect_handwriting(image_bytes, target_ink=req.target_ink)
+
+    # 3. Convert mask to PNG bytes and dilate
+    import io
+    from PIL import Image as _PILImage
+    mask_img = _PILImage.fromarray(detection.mask)
+    mask_buf = io.BytesIO()
+    mask_img.save(mask_buf, format="PNG")
+    mask_bytes = mask_buf.getvalue()
+
+    if req.dilation > 0:
+        mask_bytes = _dilate_mask(mask_bytes, iterations=req.dilation)
+
+    # 4. Inpaint
+    method_map = {
+        "telea": InpaintingMethod.OPENCV_TELEA,
+        "ns": InpaintingMethod.OPENCV_NS,
+        "auto": InpaintingMethod.AUTO,
+    }
+    inpaint_method = method_map.get(req.method, InpaintingMethod.AUTO)
+
+    result = inpaint_image(image_bytes, mask_bytes, method=inpaint_method)
+    if not result.success:
+        raise HTTPException(status_code=500, detail="Inpainting failed")
+
+    elapsed_ms = int((_time.monotonic() - t0) * 1000)
+
+    meta = {
+        "method_used": result.method_used.value if hasattr(result.method_used, "value") else str(result.method_used),
+        "handwriting_ratio": round(detection.handwriting_ratio, 4),
+        "detection_confidence": round(detection.confidence, 4),
+        "target_ink": req.target_ink,
+        "dilation": req.dilation,
+        "source_image": source,
+        "processing_time_ms": elapsed_ms,
+    }
+
+    # 5. Persist clean image (convert BGR ndarray → PNG bytes)
+    clean_png_bytes = image_to_png(result.image)
+    await update_session_db(session_id, clean_png=clean_png_bytes, handwriting_removal_meta=meta)
+
+    return {
+        **meta,
+        "image_url": f"/api/v1/ocr-pipeline/sessions/{session_id}/image/clean",
+        "session_id": session_id,
+    }

klausur-service/backend/ocr_pipeline_session_store.py

@@ -60,6 +60,13 @@ async def init_ocr_pipeline_tables():
         else:
             logger.debug("OCR pipeline tables already exist")
 
+        # Ensure new columns exist (idempotent ALTER TABLE)
+        await conn.execute("""
+            ALTER TABLE ocr_pipeline_sessions
+            ADD COLUMN IF NOT EXISTS clean_png BYTEA,
+            ADD COLUMN IF NOT EXISTS handwriting_removal_meta JSONB
+        """)
+
 
 # =============================================================================
 # SESSION CRUD
@@ -111,6 +118,7 @@ async def get_session_image(session_id: str, image_type: str) -> Optional[bytes]
         "deskewed": "deskewed_png",
         "binarized": "binarized_png",
         "dewarped": "dewarped_png",
+        "clean": "clean_png",
     }
     column = column_map.get(image_type)
     if not column:
@@ -135,11 +143,12 @@ async def update_session_db(session_id: str, **kwargs) -> Optional[Dict[str, Any
     allowed_fields = {
         'name', 'filename', 'status', 'current_step',
         'original_png', 'deskewed_png', 'binarized_png', 'dewarped_png',
+        'clean_png', 'handwriting_removal_meta',
         'deskew_result', 'dewarp_result', 'column_result', 'row_result',
         'word_result', 'ground_truth', 'auto_shear_degrees',
     }
 
-    jsonb_fields = {'deskew_result', 'dewarp_result', 'column_result', 'row_result', 'word_result', 'ground_truth'}
+    jsonb_fields = {'deskew_result', 'dewarp_result', 'column_result', 'row_result', 'word_result', 'ground_truth', 'handwriting_removal_meta'}
 
     for key, value in kwargs.items():
         if key in allowed_fields:

klausur-service/backend/services/handwriting_detection.py

@@ -6,6 +6,7 @@ Uses multiple detection methods:
 1. Color-based detection (blue/red ink)
 2. Stroke analysis (thin irregular strokes)
 3. Edge density variance
+4. Pencil detection (gray ink)
 
 DATENSCHUTZ: All processing happens locally on Mac Mini.
 """
@@ -37,12 +38,16 @@ class DetectionResult:
     detection_method: str  # Which method was primarily used
 
 
-def detect_handwriting(image_bytes: bytes) -> DetectionResult:
+def detect_handwriting(image_bytes: bytes, target_ink: str = "all") -> DetectionResult:
     """
     Detect handwriting in an image.
 
     Args:
         image_bytes: Image as bytes (PNG, JPG, etc.)
+        target_ink: Which ink types to detect:
+            - "all"     → all methods combined (incl. pencil)
+            - "colored" → only color-based (blue/red/green pen)
+            - "pencil"  → only pencil (gray ink)
 
     Returns:
         DetectionResult with binary mask where handwriting is white (255)
@@ -62,35 +67,51 @@ def detect_handwriting(image_bytes: bytes) -> DetectionResult:
 
     # Convert to BGR if needed (OpenCV format)
     if len(img_array.shape) == 2:
-        # Grayscale to BGR
         img_bgr = cv2.cvtColor(img_array, cv2.COLOR_GRAY2BGR)
     elif img_array.shape[2] == 4:
-        # RGBA to BGR
         img_bgr = cv2.cvtColor(img_array, cv2.COLOR_RGBA2BGR)
     elif img_array.shape[2] == 3:
-        # RGB to BGR
         img_bgr = cv2.cvtColor(img_array, cv2.COLOR_RGB2BGR)
     else:
         img_bgr = img_array
 
-    # Run multiple detection methods
-    color_mask, color_confidence = _detect_by_color(img_bgr)
-    stroke_mask, stroke_confidence = _detect_by_stroke_analysis(img_bgr)
-    variance_mask, variance_confidence = _detect_by_variance(img_bgr)
+    # Select detection methods based on target_ink
+    masks_and_weights = []
+
+    if target_ink in ("all", "colored"):
+        color_mask, color_conf = _detect_by_color(img_bgr)
+        masks_and_weights.append((color_mask, color_conf, "color"))
+
+    if target_ink == "all":
+        stroke_mask, stroke_conf = _detect_by_stroke_analysis(img_bgr)
+        variance_mask, variance_conf = _detect_by_variance(img_bgr)
+        masks_and_weights.append((stroke_mask, stroke_conf, "stroke"))
+        masks_and_weights.append((variance_mask, variance_conf, "variance"))
+
+    if target_ink in ("all", "pencil"):
+        pencil_mask, pencil_conf = _detect_pencil(img_bgr)
+        masks_and_weights.append((pencil_mask, pencil_conf, "pencil"))
+
+    if not masks_and_weights:
+        # Fallback: use all methods
+        color_mask, color_conf = _detect_by_color(img_bgr)
+        stroke_mask, stroke_conf = _detect_by_stroke_analysis(img_bgr)
+        variance_mask, variance_conf = _detect_by_variance(img_bgr)
+        pencil_mask, pencil_conf = _detect_pencil(img_bgr)
+        masks_and_weights = [
+            (color_mask, color_conf, "color"),
+            (stroke_mask, stroke_conf, "stroke"),
+            (variance_mask, variance_conf, "variance"),
+            (pencil_mask, pencil_conf, "pencil"),
+        ]
 
     # Combine masks using weighted average
-    weights = [color_confidence, stroke_confidence, variance_confidence]
-    total_weight = sum(weights)
+    total_weight = sum(w for _, w, _ in masks_and_weights)
 
     if total_weight > 0:
-        # Weighted combination
-        combined_mask = (
-            color_mask.astype(np.float32) * color_confidence +
-            stroke_mask.astype(np.float32) * stroke_confidence +
-            variance_mask.astype(np.float32) * variance_confidence
+        combined_mask = sum(
+            m.astype(np.float32) * w for m, w, _ in masks_and_weights
         ) / total_weight
-
-        # Threshold to binary
         combined_mask = (combined_mask > 127).astype(np.uint8) * 255
     else:
         combined_mask = np.zeros(img_bgr.shape[:2], dtype=np.uint8)
@@ -103,19 +124,11 @@ def detect_handwriting(image_bytes: bytes) -> DetectionResult:
     handwriting_pixels = np.sum(combined_mask > 0)
     handwriting_ratio = handwriting_pixels / total_pixels if total_pixels > 0 else 0
 
-    # Determine primary method
-    primary_method = "combined"
-    max_conf = max(color_confidence, stroke_confidence, variance_confidence)
-    if max_conf == color_confidence:
-        primary_method = "color"
-    elif max_conf == stroke_confidence:
-        primary_method = "stroke"
-    else:
-        primary_method = "variance"
+    # Determine primary method (highest confidence)
+    primary_method = max(masks_and_weights, key=lambda x: x[1])[2] if masks_and_weights else "combined"
+    overall_confidence = total_weight / len(masks_and_weights) if masks_and_weights else 0.0
 
-    overall_confidence = total_weight / 3.0  # Average confidence
-
-    logger.info(f"Handwriting detection: {handwriting_ratio:.2%} handwriting, "
+    logger.info(f"Handwriting detection (target_ink={target_ink}): {handwriting_ratio:.2%} handwriting, "
                 f"confidence={overall_confidence:.2f}, method={primary_method}")
 
     return DetectionResult(
@@ -180,6 +193,27 @@ def _detect_by_color(img_bgr: np.ndarray) -> Tuple[np.ndarray, float]:
     return color_mask, confidence
 
 
+def _detect_pencil(img_bgr: np.ndarray) -> Tuple[np.ndarray, float]:
+    """
+    Detect pencil marks (gray ink, ~140-220 on 255-scale).
+
+    Paper is usually >230, dark ink <130.
+    Pencil falls in the 140-220 gray range.
+    """
+    gray = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2GRAY)
+    pencil_mask = cv2.inRange(gray, 140, 220)
+
+    # Remove small noise artifacts
+    kernel = np.ones((2, 2), np.uint8)
+    pencil_mask = cv2.morphologyEx(pencil_mask, cv2.MORPH_OPEN, kernel, iterations=1)
+
+    ratio = np.sum(pencil_mask > 0) / pencil_mask.size
+    # Good confidence if pencil pixels are in a plausible range
+    confidence = 0.75 if 0.002 < ratio < 0.2 else 0.2
+
+    return pencil_mask, confidence
+
+
 def _detect_by_stroke_analysis(img_bgr: np.ndarray) -> Tuple[np.ndarray, float]:
     """
     Detect handwriting by analyzing stroke characteristics.

klausur-service/backend/services/trocr_service.py

@@ -31,8 +31,10 @@ from datetime import datetime, timedelta
 logger = logging.getLogger(__name__)
 
 # Lazy loading for heavy dependencies
-_trocr_processor = None
-_trocr_model = None
+# Cache keyed by model_name to support base and large variants simultaneously
+_trocr_models: dict = {}  # {model_name: (processor, model)}
+_trocr_processor = None  # backwards-compat alias → base-printed
+_trocr_model = None      # backwards-compat alias → base-printed
 _trocr_available = None
 _model_loaded_at = None
 
@@ -124,12 +126,14 @@ def _check_trocr_available() -> bool:
     return _trocr_available
 
 
-def get_trocr_model(handwritten: bool = False):
+def get_trocr_model(handwritten: bool = False, size: str = "base"):
     """
     Lazy load TrOCR model and processor.
 
     Args:
         handwritten: Use handwritten model instead of printed model
+        size: Model size — "base" (300 MB) or "large" (340 MB, higher accuracy
+              for exam HTR). Only applies to handwritten variant.
 
     Returns tuple of (processor, model) or (None, None) if unavailable.
     """
@@ -138,31 +142,42 @@ def get_trocr_model(handwritten: bool = False):
     if not _check_trocr_available():
         return None, None
 
-    if _trocr_processor is None or _trocr_model is None:
-        try:
-            import torch
-            from transformers import TrOCRProcessor, VisionEncoderDecoderModel
+    # Select model name
+    if size == "large" and handwritten:
+        model_name = "microsoft/trocr-large-handwritten"
+    elif handwritten:
+        model_name = "microsoft/trocr-base-handwritten"
+    else:
+        model_name = "microsoft/trocr-base-printed"
 
-            # Choose model based on use case
-            if handwritten:
-                model_name = "microsoft/trocr-base-handwritten"
-            else:
-                model_name = "microsoft/trocr-base-printed"
+    if model_name in _trocr_models:
+        return _trocr_models[model_name]
 
-            logger.info(f"Loading TrOCR model: {model_name}")
-            _trocr_processor = TrOCRProcessor.from_pretrained(model_name)
-            _trocr_model = VisionEncoderDecoderModel.from_pretrained(model_name)
+    try:
+        import torch
+        from transformers import TrOCRProcessor, VisionEncoderDecoderModel
 
-            # Use GPU if available
-            device = "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu"
-            _trocr_model.to(device)
-            logger.info(f"TrOCR model loaded on device: {device}")
+        logger.info(f"Loading TrOCR model: {model_name}")
+        processor = TrOCRProcessor.from_pretrained(model_name)
+        model = VisionEncoderDecoderModel.from_pretrained(model_name)
 
-        except Exception as e:
-            logger.error(f"Failed to load TrOCR model: {e}")
-            return None, None
+        # Use GPU if available
+        device = "cuda" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu"
+        model.to(device)
+        logger.info(f"TrOCR model loaded on device: {device}")
 
-    return _trocr_processor, _trocr_model
+        _trocr_models[model_name] = (processor, model)
+
+        # Keep backwards-compat globals pointing at base-printed
+        if model_name == "microsoft/trocr-base-printed":
+            _trocr_processor = processor
+            _trocr_model = model
+
+        return processor, model
+
+    except Exception as e:
+        logger.error(f"Failed to load TrOCR model {model_name}: {e}")
+        return None, None
 
 
 def preload_trocr_model(handwritten: bool = True) -> bool:
@@ -209,7 +224,8 @@ def get_model_status() -> Dict[str, Any]:
 async def run_trocr_ocr(
     image_data: bytes,
     handwritten: bool = False,
-    split_lines: bool = True
+    split_lines: bool = True,
+    size: str = "base",
 ) -> Tuple[Optional[str], float]:
     """
     Run TrOCR on an image.
@@ -223,11 +239,12 @@ async def run_trocr_ocr(
         image_data: Raw image bytes
         handwritten: Use handwritten model (slower but better for handwriting)
         split_lines: Whether to split image into lines first
+        size: "base" or "large" (only for handwritten variant)
 
     Returns:
         Tuple of (extracted_text, confidence)
     """
-    processor, model = get_trocr_model(handwritten=handwritten)
+    processor, model = get_trocr_model(handwritten=handwritten, size=size)
 
     if processor is None or model is None:
         logger.error("TrOCR model not available")