From b0049229d880c11198f1e0238b0c90b8faecb17d Mon Sep 17 00:00:00 2001
From: kyy <b24053@hanmaceng.co.kr>
Date: Wed, 13 Aug 2025 11:07:44 +0900
Subject: [PATCH] =?UTF-8?q?PaddleOCR=20=EC=8B=A4=ED=96=89=20=EC=A0=84=20?=
 =?UTF-8?q?=EC=9D=B4=EB=AF=B8=EC=A7=80=EB=A5=BC=203=EC=B1=84=EB=84=90=20ui?=
 =?UTF-8?q?nt8=EB=A1=9C=20=ED=91=9C=EC=A4=80=ED=99=94=ED=95=98=EC=97=AC=20?=
 =?UTF-8?q?Normalize=20IndexError=20=EB=B0=A9=EC=A7=80?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 utils/text_extractor.py | 53 +++++++++++++++++++++++++++++++++++++++--
 1 file changed, 51 insertions(+), 2 deletions(-)

diff --git a/utils/text_extractor.py b/utils/text_extractor.py
index e58ee09..d7e1a43 100644
--- a/utils/text_extractor.py
+++ b/utils/text_extractor.py
@@ -105,6 +105,41 @@ def preprocess_image_for_ocr(pil_img, page_idx=None):
     return Image.fromarray(img)
 
 
+def _to_rgb_uint8(img_np: np.ndarray) -> np.ndarray:
+    """
+    입력 이미지를 3채널 RGB, uint8 [0,255] 로 표준화
+    허용 입력: HxW, HxWx1, HxWx3, HxWx4, float[0..1]/[0..255], int 등
+    """
+    if img_np is None:
+        raise ValueError("Input image is None")
+
+    # dtype/범위 표준화
+    if img_np.dtype != np.uint8:
+        arr = img_np.astype(np.float32)
+        if arr.max() <= 1.0:  # [0,1]로 보이면 스케일업
+            arr *= 255.0
+        arr = np.clip(arr, 0, 255).astype(np.uint8)
+        img_np = arr
+
+    # 채널 표준화
+    if img_np.ndim == 2:  # HxW
+        img_np = cv2.cvtColor(img_np, cv2.COLOR_GRAY2RGB)
+    elif img_np.ndim == 3:
+        h, w, c = img_np.shape
+        if c == 1:
+            img_np = cv2.cvtColor(img_np, cv2.COLOR_GRAY2RGB)
+        elif c == 4:
+            img_np = cv2.cvtColor(img_np, cv2.COLOR_RGBA2RGB)
+        elif c == 3:
+            pass  # 그대로 사용
+        else:
+            raise ValueError(f"Unsupported channel count: {c}")
+    else:
+        raise ValueError(f"Unsupported ndim: {img_np.ndim}")
+
+    return img_np
+
+
 def extract_text_paddle_ocr(images):
     """
     PaddleOCR를 사용하여 이미지에서 텍스트 추출 및 좌표 정보 반환
@@ -121,8 +156,22 @@ def extract_text_paddle_ocr(images):
         print(f"[PaddleOCR] 페이지 {page_idx + 1} OCR로 텍스트 추출 중...")
         img_np = np.array(img)
 
-        if len(img_np.shape) == 2:  # grayscale → RGB 변환
-            img_np = cv2.cvtColor(img_np, cv2.COLOR_GRAY2RGB)
+        # ✅ 채널/타입 표준화 (grayscale/rgba/float 등 대응)
+        try:
+            img_np = _to_rgb_uint8(img_np)
+        except Exception as e:
+            print(f"[PaddleOCR] 페이지 {page_idx + 1} 입력 표준화 실패: {e}")
+            continue  # 문제 페이지 스킵 후 다음 페이지 진행
+
+        # ✅ 과도한 해상도 안정화 (최대 변 4000px)
+        h, w = img_np.shape[:2]
+        max_side = max(h, w)
+        max_side_limit = 4000
+        if max_side > max_side_limit:
+            scale = max_side_limit / max_side
+            new_size = (int(w * scale), int(h * scale))
+            img_np = cv2.resize(img_np, new_size, interpolation=cv2.INTER_AREA)
+            print(f"[PaddleOCR] Resized to {img_np.shape[1]}x{img_np.shape[0]}")
 
         results = ocr.predict(input=img_np)