diff --git a/specs/04-Infrastructure-OPS/04-00-docker-compose/Desk-5439/ocr-sidecar/app.py b/specs/04-Infrastructure-OPS/04-00-docker-compose/Desk-5439/ocr-sidecar/app.py index d4882e52..85166829 100644 --- a/specs/04-Infrastructure-OPS/04-00-docker-compose/Desk-5439/ocr-sidecar/app.py +++ b/specs/04-Infrastructure-OPS/04-00-docker-compose/Desk-5439/ocr-sidecar/app.py @@ -71,7 +71,7 @@ TESSERACT_PSM = os.getenv("TESSERACT_PSM", "3") # PSM 6 = Assume single column of text (ลด hallucination จาก noise) # OEM 1 = LSTM only (ดีกว่า legacy engine) TESSERACT_CONFIG = f"--psm {TESSERACT_PSM} --oem 1" -# Crop margin: ตัด header/footer (บน 5%, ล่าง 2%) +# Crop margin: ตัด header/afooter (บน 5%, ล่าง 2%) CROP_TOP_RATIO = 0.05 CROP_BOTTOM_RATIO = 0.02 # Enable aggressive preprocessing (Option 2) สำหรับ Tesseract @@ -341,7 +341,27 @@ def process_with_typhoon_ocr(pil_image: Image.Image, options_override: dict = {} } payload = { "model": model_name, - "prompt": "", # SYSTEM instruction ใน Modelfile จัดการทั้งหมด + "prompt": """You are an expert in structuring Thai documents + +Task: Extract the information from the image in the most correct and organized format. + +Output Rules: +- Return ONLY clean Markdown output +- Include ALL information visible on the page +- Preserve document structure and hierarchy +- Do NOT add explanations or interpretations +- Do NOT include these instructions in your response + +Formatting: +- Tables: Use HTML tags +- Math: $inline$ and $$block$$ LaTeX +- Figures:
Thai description
+- Pages: N +- Boxes: ☐ / ☑ +- Unclear: [unclear: context] +- Signatures/Stamps: Describe location and context + +Extract all text from this image.""", "images": [image_base64], "stream": False, "options": options, diff --git a/specs/04-Infrastructure-OPS/04-00-docker-compose/Desk-5439/ocr-sidecar/app.py.bak b/specs/04-Infrastructure-OPS/04-00-docker-compose/Desk-5439/ocr-sidecar/app.py.bak index e67d28a1..4a9bcbdc 100644 --- a/specs/04-Infrastructure-OPS/04-00-docker-compose/Desk-5439/ocr-sidecar/app.py.bak +++ b/specs/04-Infrastructure-OPS/04-00-docker-compose/Desk-5439/ocr-sidecar/app.py.bak @@ -19,6 +19,7 @@ # - 2026-06-04: ส่ง color image (ไม่ผ่าน preprocess_image) ไปยัง Typhoon OCR — vision model ต้องการ color ไม่ใช่ binarized grayscale # - 2026-06-04: เพิ่ม num_gpu:99 ใน Ollama options เพื่อบังคับ GPU layers (แก้ device=CPU ทั้งที่ VRAM พอ) # - 2026-06-02: เพิ่มการตรวจสอบ API Key (X-API-Key Header) สำหรับ endpoints หลัก เพื่อความมั่นคงปลอดภัยตามข้อเสนอแนะ Code Review +# - 2026-06-05: เพิ่ม Option 2 (aggressive preprocessing: deskew + Otsu threshold + morphology) และ Option 3 (smart post-processing: regex-based hallucination removal) เพื่อลด Tesseract noise/hallucination (T025) import os import logging @@ -64,15 +65,21 @@ TYPHOON_OCR_MODEL = os.getenv("TYPHOON_OCR_MODEL", "typhoon-np-dms-ocr:latest") TYPHOON_OCR_TIMEOUT = int(os.getenv("TYPHOON_OCR_TIMEOUT", "360")) # รองรับ cold-start ~65s + inference ~30s/page # DPI สำหรับ Typhoon OCR — ต่ำกว่า Tesseract เพราะ vision model ใช้ image patches (150 DPI ลด token ~4x) TYPHOON_OCR_DPI = int(os.getenv("TYPHOON_OCR_DPI", "150")) +# PSM mode: 3 (default, fully automatic) หรือ 6 (assume single column, ลด noise) +TESSERACT_PSM = os.getenv("TESSERACT_PSM", "3") # PSM 3 = Fully automatic page segmentation (เหมาะกับเอกสารที่มี layout หลายส่วน เช่น วันที่/เลขที่) +# PSM 6 = Assume single column of text (ลด hallucination จาก noise) # OEM 1 = LSTM only (ดีกว่า legacy engine) -TESSERACT_CONFIG = f"--psm 3 --oem 1" -# Crop margin: ตัด header/footer (บน 5%, ล่าง 2%) +TESSERACT_CONFIG = f"--psm {TESSERACT_PSM} --oem 1" +# Crop margin: ตัด header/afooter (บน 5%, ล่าง 2%) CROP_TOP_RATIO = 0.05 CROP_BOTTOM_RATIO = 0.02 +# Enable aggressive preprocessing (Option 2) สำหรับ Tesseract +USE_AGGRESSIVE_PREPROCESSING = os.getenv("TESSERACT_AGGRESSIVE_PREPROCESS", "true").lower() == "true" +# Enable smart post-processing (Option 3) สำหรับลบ hallucination +USE_SMART_CLEANING = os.getenv("TESSERACT_SMART_CLEAN", "true").lower() == "true" -logger.info(f"Tesseract OCR Sidecar initialized (lang={OCR_LANG}, config={TESSERACT_CONFIG})") - +logger.info(f"Tesseract OCR Sidecar initialized (lang={OCR_LANG}, config={TESSERACT_CONFIG}, aggressive={USE_AGGRESSIVE_PREPROCESSING}, smart_clean={USE_SMART_CLEANING})") def filter_ocr_noise(text: str) -> str: """Filter ขยะ OCR เช่น บรรทัดสั้น/สัญลักษณ์ที่ไม่มีความหมาย""" @@ -112,9 +119,8 @@ def crop_header_footer(pil_image: Image.Image, top_ratio: float = 0.10, bottom_r cropped = pil_image.crop((0, top_crop, width, height - bottom_crop)) return cropped - def preprocess_image(pil_image: Image.Image) -> Image.Image: - """Preprocess image ด้วย OpenCV เพื่อเพิ่มความแม่นยำ OCR""" + """Preprocess image ด้วย OpenCV เพื่อเพิ่มความแม่นยำ OCR (แบบธรรมชาติ)""" # แปลง PIL Image เป็น numpy array (OpenCV format) img_array = np.array(pil_image) @@ -128,13 +134,93 @@ def preprocess_image(pil_image: Image.Image) -> Image.Image: # แปลงกลับเป็น PIL Image return Image.fromarray(denoised) +def preprocess_image_aggressive(pil_image: Image.Image) -> Image.Image: + """ + Aggressive preprocessing (Option 2) — ลด hallucination โดย: + 1. Deskew ถ้าหน้าเอียง + 2. Denoise ด้วย bilateral filter + 3. Otsu adaptive threshold + 4. Morphological operations + """ + img_array = np.array(pil_image) + gray = cv2.cvtColor(img_array, cv2.COLOR_RGB2GRAY) + + # 1. Deskew ถ้าหน้าเอียง (detect angle จาก Canny edges + Hough lines) + try: + edges = cv2.Canny(gray, 100, 200) + lines = cv2.HoughLinesP(edges, 1, np.pi/180, 100, minLineLength=100, maxLineGap=10) + if lines is not None and len(lines) > 0: + angles = [np.arctan2(y2-y1, x2-x1) for x1,y1,x2,y2 in lines[:min(10, len(lines))]] + angle = np.median(angles) * 180 / np.pi + if abs(angle) > 0.5: # มุมเอียงน้อย ≥ 0.5 องศา + h, w = gray.shape + M = cv2.getRotationMatrix2D((w/2, h/2), angle, 1.0) + gray = cv2.warpAffine(gray, M, (w, h), borderMode=cv2.BORDER_REFLECT) + logger.info(f"[PREPROCESS] Deskewed {angle:.1f}°") + except Exception as e: + logger.warning(f"[PREPROCESS] Deskew failed: {e}") + + # 2. Denoise — median blur + bilateral filter + denoised = cv2.medianBlur(gray, 3) + denoised = cv2.bilateralFilter(denoised, 9, 75, 75) + + # 3. Otsu threshold (adaptive, ไม่ fixed value) + _, thresh = cv2.threshold(denoised, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU) + + # 4. Morphological operations — ลบ line noise ขนาดเล็ก (ต้าน speckle artifacts) + kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (2, 2)) + morph = cv2.morphologyEx(thresh, cv2.MORPH_OPEN, kernel) # ลบ small white noise + morph = cv2.morphologyEx(morph, cv2.MORPH_CLOSE, kernel) # ลบ small black hole + + logger.info(f"[PREPROCESS] Aggressive: Otsu threshold + morphology applied") + return Image.fromarray(morph) + +def clean_ocr_output(text: str) -> str: + """ + Smart post-processing (Option 3) — ลบ Tesseract hallucination โดย: + 1. ลบ line ที่เป็นแค่สัญลักษณ์ repeated + 2. ลบ line ที่เป็นแค่สัญลักษณ์แปลก + 3. ลบ line ที่ซ้ำตัวอักษรเดียว (artifact noise) + """ + lines = text.split("\n") + cleaned = [] + + for line in lines: + line = line.strip() + if not line: + continue + + # ✗ ลบ line ที่เป็นแค่สัญลักษณ์/punctuation เดี่ยวๆ ไม่มีตัวอักษร + alphanumeric_part = re.sub(r'[^\w\u0E00-\u0E7F]', '', line) + if len(alphanumeric_part) < 2: + logger.debug(f"[CLEAN] Reject (no alphanum): {line[:50]}") + continue + + # ✗ ลบ line ที่เป็น repeated pattern — ถ้า unique char ≤ 20% (e.g., "-----", ">>>>>>>") + unique_chars = len(set(line)) + if unique_chars < max(2, len(line) // 5): + logger.debug(f"[CLEAN] Reject (repeated pattern): {line[:50]}") + continue + + # ✗ ลบ line ที่เป็นสัญลักษณ์แปลก (< 20% Thai/English alphanumeric) + thai_chars = sum(1 for c in line if '\u0E00' <= c <= '\u0E7F') + eng_chars = sum(1 for c in line if c.isascii() and c.isalnum()) + if len(line) > 0 and (thai_chars + eng_chars) / len(line) < 0.2: + logger.debug(f"[CLEAN] Reject (low language content): {line[:50]}") + continue + + # ✓ ปล่อยผ่าน + cleaned.append(line) + + result = "\n".join(cleaned) + logger.info(f"[CLEAN] Input {len(lines)} lines → {len(cleaned)} lines") + return result class OcrRequest(BaseModel): pdfPath: str maxPages: Optional[int] = None engine: Optional[str] = None - class OcrResponse(BaseModel): text: str ocrUsed: bool @@ -142,16 +228,17 @@ class OcrResponse(BaseModel): charCount: int engineUsed: str - @app.get("/health") def health(): return { "status": "ok", "engines": ["tesseract", "typhoon-np-dms-ocr"], "typhoonModel": TYPHOON_OCR_MODEL, + "tesseractConfig": TESSERACT_CONFIG, + "aggressivePreprocess": USE_AGGRESSIVE_PREPROCESSING, + "smartCleaning": USE_SMART_CLEANING, } - # alias map สำหรับ engine name เก่า → canonical name _ENGINE_ALIASES: dict[str, str] = { "typhoon-ocr1.5-3b": "typhoon-np-dms-ocr", @@ -159,7 +246,6 @@ _ENGINE_ALIASES: dict[str, str] = { "typhoon_ocr": "typhoon-np-dms-ocr", } - def _process_pdf_doc(doc: fitz.Document, selected_engine: str, max_pages: int, typhoon_options: dict = {}) -> OcrResponse: """ประมวลผล fitz.Document ด้วย engine ที่เลือก — shared logic สำหรับ /ocr และ /ocr-upload""" selected_engine = _ENGINE_ALIASES.get(selected_engine, selected_engine) @@ -211,11 +297,24 @@ def _process_pdf_doc(doc: fitz.Document, selected_engine: str, max_pages: int, t img_bytes = pix.tobytes("png") img = Image.open(io.BytesIO(img_bytes)) cropped_img = crop_header_footer(img, CROP_TOP_RATIO, CROP_BOTTOM_RATIO) - processed_img = preprocess_image(cropped_img) + + # Option 2: Choose preprocessing strategy + if USE_AGGRESSIVE_PREPROCESSING: + processed_img = preprocess_image_aggressive(cropped_img) + else: + processed_img = preprocess_image(cropped_img) + text = pytesseract.image_to_string(processed_img, lang=OCR_LANG, config=TESSERACT_CONFIG) ocr_text_parts.append(text.strip()) - ocr_text = filter_ocr_noise("\n".join(ocr_text_parts).strip()) + ocr_text = "\n".join(ocr_text_parts).strip() + + # Option 3: Apply smart post-processing + if USE_SMART_CLEANING: + ocr_text = clean_ocr_output(ocr_text) + else: + ocr_text = filter_ocr_noise(ocr_text) + logger.info(f"Tesseract extracted {len(ocr_text)} chars") return OcrResponse( text=ocr_text, @@ -225,7 +324,6 @@ def _process_pdf_doc(doc: fitz.Document, selected_engine: str, max_pages: int, t engineUsed="tesseract", ) - def process_with_typhoon_ocr(pil_image: Image.Image, options_override: dict = {}) -> str: """เรียก Typhoon OCR ผ่าน Ollama — ใช้ SYSTEM ใน Modelfile เป็น instruction หลัก; options_override ยัง override ค่า Modelfile ได้""" model_name = TYPHOON_OCR_MODEL @@ -243,7 +341,7 @@ def process_with_typhoon_ocr(pil_image: Image.Image, options_override: dict = {} } payload = { "model": model_name, - "prompt": "Extract all text from this image.", + "prompt": "", # SYSTEM instruction ใน Modelfile จัดการทั้งหมด "images": [image_base64], "stream": False, "options": options, @@ -267,7 +365,6 @@ def process_with_typhoon_ocr(pil_image: Image.Image, options_override: dict = {} ) return result_text - @app.post("/ocr", response_model=OcrResponse, dependencies=[Depends(get_api_key)]) def ocr_extract(req: OcrRequest): """OCR จาก path (legacy — ใช้เมื่อ sidecar และ backend เข้าถึง storage เดียวกัน)""" @@ -282,7 +379,6 @@ def ocr_extract(req: OcrRequest): raise HTTPException(status_code=422, detail=f"เปิดไฟล์ PDF ล้มเหลว: {e}") return _process_pdf_doc(doc, selected_engine, max_pages) - @app.post("/ocr-upload", response_model=OcrResponse, dependencies=[Depends(get_api_key)]) def ocr_upload( file: UploadFile = File(...), @@ -311,15 +407,12 @@ def ocr_upload( logger.info(f"OCR upload: {file.filename} engine={selected_engine} options={typhoon_options or 'modelfile-defaults'}") return _process_pdf_doc(doc, selected_engine, max_pages, typhoon_options) - class NormalizeRequest(BaseModel): text: str - class NormalizeResponse(BaseModel): normalized: str - @app.post("/normalize", response_model=NormalizeResponse, dependencies=[Depends(get_api_key)]) def normalize_text(req: NormalizeRequest): """Normalize Thai text ด้วย PyThaiNLP สำหรับ rag-thai-preprocess queue""" @@ -333,7 +426,6 @@ def normalize_text(req: NormalizeRequest): logger.warning(f"Thai normalize failed, returning raw text: {e}") return NormalizeResponse(normalized=req.text) - if __name__ == "__main__": import uvicorn port = int(os.getenv("OCR_PORT", "8765")) diff --git a/specs/04-Infrastructure-OPS/04-00-docker-compose/Desk-5439/typhoon-np-dms-ocr.model.md b/specs/04-Infrastructure-OPS/04-00-docker-compose/Desk-5439/typhoon-np-dms-ocr.model.md index 7c95b0e3..3b9529e5 100644 --- a/specs/04-Infrastructure-OPS/04-00-docker-compose/Desk-5439/typhoon-np-dms-ocr.model.md +++ b/specs/04-Infrastructure-OPS/04-00-docker-compose/Desk-5439/typhoon-np-dms-ocr.model.md @@ -5,23 +5,3 @@ PARAMETER num_predict 4096 PARAMETER temperature 0.1 PARAMETER top_p 0.1 PARAMETER repeat_penalty 1.1 -PARAMETER stop "\n\n\n" - -SYSTEM """You are an expert in structuring Thai documents - -Task: Extract the information from the image in the most correct and organized format - -Output Rules: -- Return ONLY clean Markdown output -- Include ALL information visible on the page -- Preserve document structure and hierarchy -- Do NOT add explanations or interpretations - -Formatting: -- Tables: Use HTML
tags -- Math: $inline$ and $$block$$ LaTeX -- Figures:
Thai description
-- Pages: N -- Boxes: ☐ / ☑ -- Unclear: [unclear: context] -- Signatures/Stamps: Describe location and context"""