Configure PaddleOCR to use Korean PP-OCRv5

app.py

@@ -9,7 +9,7 @@ import spaces
 import torch
 from PIL import Image
 from transformers import AutoTokenizer, AutoModelForCausalLM
-from huggingface_hub import login
+from huggingface_hub import login, snapshot_download
 from paddleocr import PaddleOCR
 
 # Hugging Face 토큰으로 로그인 (Spaces Secret에서 가져옴)
@@ -24,14 +24,30 @@ MED_MODEL_ID = "google/gemma-2-2b-it"
 OCR_READER = None
 MED_MODEL = None
 MED_TOKENIZER = None
+OCR_MODEL_REPO_ID = "PaddlePaddle/korean_PP-OCRv5_mobile_rec"
 
 def load_models():
     """모델들을 한 번만 로드"""
     global OCR_READER, MED_MODEL, MED_TOKENIZER
 
     if OCR_READER is None:
-        print("🔄 Loading PaddleOCR (Korean)...")
-        OCR_READER = PaddleOCR(lang='korean', use_angle_cls=True, show_log=False)
+        print("🔄 Loading PaddleOCR (Korean PP-OCRv5 mobile recognition)...")
+        rec_model_dir = snapshot_download(
+            OCR_MODEL_REPO_ID,
+            allow_patterns=[
+                "*.pdmodel",
+                "*.pdiparams",
+                "*.pdparams",
+                "*.json",
+                "*.yml",
+            ],
+        )
+        OCR_READER = PaddleOCR(
+            lang='korean',
+            use_angle_cls=True,
+            show_log=False,
+            rec_model_dir=rec_model_dir,
+        )
         print("✅ PaddleOCR loaded!")
 
     if MED_MODEL is None: