Update app.py

app.py

@@ -1,35 +1,42 @@
 import os
-
-# ✅ 修正權限錯誤：改用 /tmp 資料夾儲存 cache
-os.environ["TRANSFORMERS_CACHE"] = "/tmp/.cache"
-os.environ["HF_HOME"] = "/tmp/.cache"
-os.environ["TORCH_HOME"] = "/tmp/.cache"
-os.environ["HF_DATASETS_CACHE"] = "/tmp/.cache"
+import io
+import json
+import requests
+import torch
+import pytz
+import pytesseract
+import cv2
+import numpy as np
+from PIL import Image
+from datetime import datetime
+from typing import Optional, List
 
 from fastapi import FastAPI, HTTPException, File, UploadFile
 from fastapi.middleware.cors import CORSMiddleware
 from fastapi.staticfiles import StaticFiles
 from fastapi.responses import FileResponse, JSONResponse
 from pydantic import BaseModel
-from datetime import datetime
-from typing import Optional, List
+
 from firebase_admin import credentials, firestore
 import firebase_admin
-import pytz
-import json
-import requests
-import torch
-import pytesseract
-import cv2
-import numpy as np
-from PIL import Image
-import io
+
 from AI_Model_architecture import BertLSTM_CNN_Classifier
 from bert_explainer import analyze_text as bert_analyze_text
 
+# ─────────────────────────────────────────────────────────────────────────────
+# 0. 解決 Cache 權限問題：將各大 Cache 資料夾都指向 /tmp/.cache
+os.environ["TRANSFORMERS_CACHE"] = "/tmp/.cache"
+os.environ["HF_HOME"] = "/tmp/.cache"
+os.environ["TORCH_HOME"] = "/tmp/.cache"
+os.environ["HF_DATASETS_CACHE"] = "/tmp/.cache"
+
+# 1. 指定 Tesseract OCR 執行檔路徑（Hugging Face Space 預設已安裝 tesseract-ocr）
+pytesseract.pytesseract.tesseract_cmd = "/usr/bin/tesseract"
+# ─────────────────────────────────────────────────────────────────────────────
+
 app = FastAPI(
     title="詐騙訊息辨識 API",
-    description="使用 BERT 模型分析輸入文字是否為詐騙內容",
+    description="使用 BERT 模型與 OCR 圖像前處理，辨識文字並做詐騙判斷",
     version="1.0.0"
 )
 
@@ -41,22 +48,32 @@ app.add_middleware(
     allow_headers=["*"],
 )
 
+# 掛載根目錄為靜態檔，用於提供 index.html
 app.mount("/static", StaticFiles(directory="."), name="static")
 
 @app.get("/", response_class=FileResponse)
 async def serve_index():
+    """
+    回傳根目錄的 index.html
+    """
     return FileResponse("index.html")
 
+# ─────────────────────────────────────────────────────────────────────────────
+# 2. Firebase 初始化（以環境變數 FIREBASE_CREDENTIALS 儲存 service account JSON 字串）
 try:
     cred_data = os.getenv("FIREBASE_CREDENTIALS")
     if not cred_data:
         raise ValueError("FIREBASE_CREDENTIALS 環境變數未設置")
-    cred = credentials.Certificate({"type": "service_account", **json.loads(cred_data)})
-    firebase_admin.initialize_app(cred)
+    firebase_cred = credentials.Certificate({"type": "service_account", **json.loads(cred_data)})
+    firebase_admin.initialize_app(firebase_cred)
     db = firestore.client()
 except Exception as e:
+    # 若初始化失敗，印在 Console，但不讓整個 app 崩潰
     print(f"Firebase 初始化錯誤: {e}")
+# ─────────────────────────────────────────────────────────────────────────────
 
+# ─────────────────────────────────────────────────────────────────────────────
+# 3. 下載並載入 PyTorch BERT+LSTM+CNN 模型
 model_path = "/tmp/model.pth"
 model_url = "https://huggingface.co/jerrynnms/scam-model/resolve/main/model.pth"
 if not os.path.exists(model_path):
@@ -65,12 +82,15 @@ if not os.path.exists(model_path):
         with open(model_path, "wb") as f:
             f.write(response.content)
     else:
-        raise FileNotFoundError("❌ 無法從 Hugging Face 載入 model.pth")
+        raise FileNotFoundError("❌ 無法從 Hugging Face 下載 model.pth")
 
 model = BertLSTM_CNN_Classifier()
 model.load_state_dict(torch.load(model_path, map_location="cpu"))
 model.eval()
+# ─────────────────────────────────────────────────────────────────────────────
 
+# ─────────────────────────────────────────────────────────────────────────────
+# 4. 定義 Pydantic Request / Response Model
 class TextAnalysisRequest(BaseModel):
     text: str
     user_id: Optional[str] = None
@@ -81,9 +101,13 @@ class TextAnalysisResponse(BaseModel):
     suspicious_keywords: List[str]
     analysis_timestamp: datetime
     text_id: str
+# ─────────────────────────────────────────────────────────────────────────────
 
 @app.post("/predict", response_model=TextAnalysisResponse)
 async def analyze_text_api(request: TextAnalysisRequest):
+    """
+    文字輸入分析：回傳是否為詐騙訊息、信心度、可疑關鍵詞清單
+    """
     try:
         tz = pytz.timezone("Asia/Taipei")
         now = datetime.now(tz)
@@ -91,8 +115,10 @@ async def analyze_text_api(request: TextAnalysisRequest):
         date_str = now.strftime("%Y-%m-%d %H:%M:%S")
         collection = now.strftime("%Y%m%d")
 
+        # 使用 Bert+LSTM+CNN 模型做文字判斷
         result = bert_analyze_text(request.text)
 
+        # 把結果存到 Firestore
         record = {
             "text_id": doc_id,
             "text": request.text,
@@ -101,7 +127,6 @@ async def analyze_text_api(request: TextAnalysisRequest):
             "timestamp": date_str,
             "type": "text_analysis"
         }
-
         db.collection(collection).document(doc_id).set(record)
 
         return TextAnalysisResponse(
@@ -116,6 +141,9 @@ async def analyze_text_api(request: TextAnalysisRequest):
 
 @app.post("/feedback")
 async def save_user_feedback(feedback: dict):
+    """
+    使用者回饋：把自訂的 feedback JSON 存到 Firestore 的 user_feedback collection
+    """
     try:
         tz = pytz.timezone("Asia/Taipei")
         timestamp_str = datetime.now(tz).strftime("%Y-%m-%d %H:%M:%S")
@@ -126,23 +154,79 @@ async def save_user_feedback(feedback: dict):
     except Exception as e:
         raise HTTPException(status_code=500, detail=str(e))
 
-# ✅ 加入 OpenCV 圖像前處理函數
-def preprocess_image_for_ocr(pil_image):
-    img = np.array(pil_image.convert('RGB'))
-    gray = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)
-    _, thresh = cv2.threshold(gray, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
-    scaled = cv2.resize(thresh, None, fx=1.5, fy=1.5, interpolation=cv2.INTER_LINEAR)
-    return Image.fromarray(scaled)
+# ─────────────────────────────────────────────────────────────────────────────
+# 5. OCR 前處理：灰階 → 中值去噪 → 自適應二值化 → 形態學閉運算 → 校正傾斜 → 放大 & 平滑
+def preprocess_image_for_ocr(pil_image: Image.Image) -> Image.Image:
+    """
+    完整前處理邏輯：
+    1. PIL Image (RGB) → NumPy (BGR)
+    2. 轉灰階
+    3. 中值去噪 (MedianBlur)
+    4. 自適應二值化 (Adaptive Threshold)
+    5. 形態學閉運算 (Morphological Close)
+    6. 校正傾斜 (Deskew)
+    7. 放大兩倍 & GaussianBlur 平滑
+    8. NumPy → PIL 回傳
+    """
+    # 1. PIL → NumPy (RGB -> BGR)
+    img = np.array(pil_image.convert("RGB"))[:, :, ::-1]
+    # 2. 轉灰階
+    gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
+    # 3. 中值去噪
+    denoised = cv2.medianBlur(gray, 3)
+    # 4. 自適應二值化
+    thresh = cv2.adaptiveThreshold(
+        denoised, 255,
+        cv2.ADAPTIVE_THRESH_GAUSSIAN_C,
+        cv2.THRESH_BINARY,
+        11, 2
+    )
+    # 5. 形態學閉運算 (kernel=2x2)
+    kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (2, 2))
+    morph = cv2.morphologyEx(thresh, cv2.MORPH_CLOSE, kernel, iterations=1)
+    # 6. 校正傾斜 (Deskew)
+    coords = np.column_stack(np.where(morph > 0))
+    if coords.shape[0] > 0:
+        angle = cv2.minAreaRect(coords)[-1]
+        if angle < -45:
+            angle = -(90 + angle)
+        else:
+            angle = -angle
+        (h, w) = morph.shape
+        M = cv2.getRotationMatrix2D((w // 2, h // 2), angle, 1.0)
+        morph = cv2.warpAffine(
+            morph, M, (w, h),
+            flags=cv2.INTER_CUBIC,
+            borderMode=cv2.BORDER_REPLICATE
+        )
+    # 7. 放大兩倍 & GaussianBlur 平滑
+    scaled = cv2.resize(morph, None, fx=2.0, fy=2.0, interpolation=cv2.INTER_CUBIC)
+    smoothed = cv2.GaussianBlur(scaled, (3, 3), 0)
+    # 8. NumPy → PIL
+    return Image.fromarray(smoothed)
+# ─────────────────────────────────────────────────────────────────────────────
 
 @app.post("/analyze-image")
 async def analyze_uploaded_image(file: UploadFile = File(...)):
+    """
+    圖片上傳並進行 OCR 辨識，擷取文字後再用 BERT 模型做詐騙分析
+    """
     try:
+        # 1. 讀取上傳的檔案 bytes
         image_bytes = await file.read()
         image = Image.open(io.BytesIO(image_bytes))
 
+        # 2. 對 PIL Image 做完整前處理
         processed_image = preprocess_image_for_ocr(image)
-        extracted_text = pytesseract.image_to_string(processed_image, lang="chi_tra+eng").strip()
 
+        # 3. 帶參數呼叫 pytesseract OCR
+        custom_config = r"-l chi_tra+eng --oem 3 --psm 6"
+        extracted_text = pytesseract.image_to_string(
+            processed_image,
+            config=custom_config
+        ).strip()
+
+        # 如果 OCR 完全抓不到任何文字，就回傳「無法辨識」
         if not extracted_text:
             return {
                 "extracted_text": "",
@@ -153,6 +237,7 @@ async def analyze_uploaded_image(file: UploadFile = File(...)):
                 }
             }
 
+        # 4. 如果擷取到文字，就套用 BERT 模型做詐騙分析
         result = bert_analyze_text(extracted_text)
 
         return {
@@ -161,4 +246,12 @@ async def analyze_uploaded_image(file: UploadFile = File(...)):
         }
 
     except Exception as e:
-        raise HTTPException(status_code=500, detail=f"圖片辨識錯誤：{str(e)}")
+        # 任何錯誤都以 500 回傳
+        raise HTTPException(status_code=500, detail=f"圖片辨識錯誤：{str(e)}")
+
+# ─────────────────────────────────────────────────────────────────────────────
+# 6. 啟動程式入口：讓本機或 Hugging Face Space 都能用 uvicorn 直接執行
+if __name__ == "__main__":
+    import uvicorn
+    port = int(os.environ.get("PORT", 7860))
+    uvicorn.run(app, host="0.0.0.0", port=port)