Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -27,7 +27,7 @@ from typing import Optional, List
|
|
| 27 |
from fastapi import FastAPI, HTTPException, File, UploadFile
|
| 28 |
from fastapi.middleware.cors import CORSMiddleware
|
| 29 |
from fastapi.staticfiles import StaticFiles
|
| 30 |
-
from fastapi.responses import FileResponse
|
| 31 |
from pydantic import BaseModel
|
| 32 |
|
| 33 |
from firebase_admin import credentials, firestore
|
|
@@ -74,7 +74,7 @@ except Exception as e:
|
|
| 74 |
|
| 75 |
|
| 76 |
# ──────────────────────────────────────────────────────────────────────────
|
| 77 |
-
# 載入模型
|
| 78 |
model_path = "/tmp/model.pth"
|
| 79 |
model_url = "https://huggingface.co/jerrynnms/scam-model/resolve/main/model.pth"
|
| 80 |
if not os.path.exists(model_path):
|
|
@@ -82,6 +82,7 @@ if not os.path.exists(model_path):
|
|
| 82 |
if response.status_code == 200:
|
| 83 |
with open(model_path, "wb") as f:
|
| 84 |
f.write(response.content)
|
|
|
|
| 85 |
else:
|
| 86 |
raise FileNotFoundError("❌ 無法從 Hugging Face 下載 model.pth")
|
| 87 |
|
|
@@ -108,6 +109,9 @@ class TextAnalysisResponse(BaseModel):
|
|
| 108 |
|
| 109 |
@app.post("/predict", response_model=TextAnalysisResponse)
|
| 110 |
async def analyze_text_api(request: TextAnalysisRequest):
|
|
|
|
|
|
|
|
|
|
| 111 |
try:
|
| 112 |
tz = pytz.timezone("Asia/Taipei")
|
| 113 |
now = datetime.now(tz)
|
|
@@ -127,7 +131,8 @@ async def analyze_text_api(request: TextAnalysisRequest):
|
|
| 127 |
}
|
| 128 |
try:
|
| 129 |
db.collection(collection).document(doc_id).set(record)
|
| 130 |
-
except:
|
|
|
|
| 131 |
pass
|
| 132 |
|
| 133 |
return TextAnalysisResponse(
|
|
@@ -143,6 +148,9 @@ async def analyze_text_api(request: TextAnalysisRequest):
|
|
| 143 |
|
| 144 |
@app.post("/feedback")
|
| 145 |
async def save_user_feedback(feedback: dict):
|
|
|
|
|
|
|
|
|
|
| 146 |
try:
|
| 147 |
tz = pytz.timezone("Asia/Taipei")
|
| 148 |
timestamp_str = datetime.now(tz).strftime("%Y-%m-%d %H:%M:%S")
|
|
@@ -150,7 +158,7 @@ async def save_user_feedback(feedback: dict):
|
|
| 150 |
feedback["used_in_training"] = False
|
| 151 |
try:
|
| 152 |
db.collection("user_feedback").add(feedback)
|
| 153 |
-
except:
|
| 154 |
pass
|
| 155 |
return {"message": "✅ 已記錄使用者回饋"}
|
| 156 |
except Exception as e:
|
|
@@ -160,41 +168,48 @@ async def save_user_feedback(feedback: dict):
|
|
| 160 |
# ──────────────────────────────────────────────────────────────────────────
|
| 161 |
# 強化 OCR 前處理 + 附帶 Debug 圖輸出
|
| 162 |
def preprocess_image_for_ocr(pil_image: Image.Image) -> Image.Image:
|
| 163 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 164 |
img = np.array(pil_image.convert("RGB"))[:, :, ::-1]
|
| 165 |
|
| 166 |
-
# 灰階 + CLAHE
|
| 167 |
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
|
| 168 |
clahe = cv2.createCLAHE(clipLimit=3.0, tileGridSize=(8, 8))
|
| 169 |
enhanced = clahe.apply(gray)
|
| 170 |
-
|
| 171 |
-
# Save debug file: CLAHE 增強後的灰階
|
| 172 |
Image.fromarray(enhanced).save("/tmp/debug_clahe.png")
|
| 173 |
|
| 174 |
-
# HSV 色彩分離 (過濾橘色背景)
|
| 175 |
hsv = cv2.cvtColor(img, cv2.COLOR_BGR2HSV)
|
| 176 |
lower_orange = np.array([5, 100, 100])
|
| 177 |
upper_orange = np.array([20, 255, 255])
|
| 178 |
mask_orange = cv2.inRange(hsv, lower_orange, upper_orange)
|
| 179 |
-
#
|
| 180 |
Image.fromarray(mask_orange).save("/tmp/debug_mask_orange.png")
|
| 181 |
|
|
|
|
| 182 |
filtered = enhanced.copy()
|
| 183 |
filtered[mask_orange > 0] = 255
|
| 184 |
-
|
| 185 |
-
# Save debug file: 過濾橘色後的灰階
|
| 186 |
Image.fromarray(filtered).save("/tmp/debug_filtered.png")
|
| 187 |
|
| 188 |
-
# 固定閾值反向二值化
|
| 189 |
_, thresh = cv2.threshold(filtered, 200, 255, cv2.THRESH_BINARY_INV)
|
| 190 |
-
#
|
| 191 |
Image.fromarray(thresh).save("/tmp/debug_thresh.png")
|
| 192 |
|
| 193 |
-
# 放大3倍 & GaussianBlur
|
| 194 |
scaled = cv2.resize(thresh, None, fx=3.0, fy=3.0, interpolation=cv2.INTER_CUBIC)
|
| 195 |
smoothed = cv2.GaussianBlur(scaled, (5, 5), 0)
|
| 196 |
-
|
| 197 |
-
# Save final debug processed image
|
| 198 |
Image.fromarray(smoothed).save("/tmp/debug_processed.png")
|
| 199 |
|
| 200 |
return Image.fromarray(smoothed)
|
|
@@ -204,38 +219,56 @@ def preprocess_image_for_ocr(pil_image: Image.Image) -> Image.Image:
|
|
| 204 |
@app.post("/analyze-image")
|
| 205 |
async def analyze_uploaded_image(file: UploadFile = File(...)):
|
| 206 |
"""
|
| 207 |
-
圖片上傳並進行 OCR 辨識
|
| 208 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 209 |
"""
|
|
|
|
|
|
|
|
|
|
| 210 |
try:
|
|
|
|
| 211 |
image_bytes = await file.read()
|
| 212 |
-
|
|
|
|
|
|
|
| 213 |
|
| 214 |
-
# 強化前處理 (並
|
| 215 |
-
processed_image = preprocess_image_for_ocr(
|
| 216 |
|
|
|
|
| 217 |
custom_config = r"-l chi_tra+eng --oem 3 --psm 6"
|
| 218 |
extracted_text = pytesseract.image_to_string(
|
| 219 |
processed_image,
|
| 220 |
config=custom_config
|
| 221 |
).strip()
|
|
|
|
| 222 |
|
|
|
|
| 223 |
if not extracted_text:
|
| 224 |
-
return {
|
| 225 |
"extracted_text": "",
|
| 226 |
"analysis_result": {
|
| 227 |
"status": "無法辨識",
|
| 228 |
"confidence": 0.0,
|
| 229 |
"suspicious_keywords": ["無法擷取分析結果"]
|
| 230 |
}
|
| 231 |
-
}
|
| 232 |
|
|
|
|
| 233 |
result = bert_analyze_text(extracted_text)
|
| 234 |
-
return {
|
| 235 |
"extracted_text": extracted_text,
|
| 236 |
"analysis_result": result
|
| 237 |
-
}
|
|
|
|
| 238 |
except Exception as e:
|
|
|
|
|
|
|
|
|
|
| 239 |
raise HTTPException(status_code=500, detail=f"圖片辨識失敗:{str(e)}")
|
| 240 |
|
| 241 |
|
|
|
|
| 27 |
from fastapi import FastAPI, HTTPException, File, UploadFile
|
| 28 |
from fastapi.middleware.cors import CORSMiddleware
|
| 29 |
from fastapi.staticfiles import StaticFiles
|
| 30 |
+
from fastapi.responses import FileResponse, JSONResponse
|
| 31 |
from pydantic import BaseModel
|
| 32 |
|
| 33 |
from firebase_admin import credentials, firestore
|
|
|
|
| 74 |
|
| 75 |
|
| 76 |
# ──────────────────────────────────────────────────────────────────────────
|
| 77 |
+
# 載入 BERT+LSTM+CNN 模型
|
| 78 |
model_path = "/tmp/model.pth"
|
| 79 |
model_url = "https://huggingface.co/jerrynnms/scam-model/resolve/main/model.pth"
|
| 80 |
if not os.path.exists(model_path):
|
|
|
|
| 82 |
if response.status_code == 200:
|
| 83 |
with open(model_path, "wb") as f:
|
| 84 |
f.write(response.content)
|
| 85 |
+
print("✅ 模型下載完成")
|
| 86 |
else:
|
| 87 |
raise FileNotFoundError("❌ 無法從 Hugging Face 下載 model.pth")
|
| 88 |
|
|
|
|
| 109 |
|
| 110 |
@app.post("/predict", response_model=TextAnalysisResponse)
|
| 111 |
async def analyze_text_api(request: TextAnalysisRequest):
|
| 112 |
+
"""
|
| 113 |
+
純文字輸入分析:使用 BERT 模型判斷詐騙與否,並取得可疑關鍵詞
|
| 114 |
+
"""
|
| 115 |
try:
|
| 116 |
tz = pytz.timezone("Asia/Taipei")
|
| 117 |
now = datetime.now(tz)
|
|
|
|
| 131 |
}
|
| 132 |
try:
|
| 133 |
db.collection(collection).document(doc_id).set(record)
|
| 134 |
+
except Exception:
|
| 135 |
+
# 如果 Firestore 無法寫入,也不影響回傳結果
|
| 136 |
pass
|
| 137 |
|
| 138 |
return TextAnalysisResponse(
|
|
|
|
| 148 |
|
| 149 |
@app.post("/feedback")
|
| 150 |
async def save_user_feedback(feedback: dict):
|
| 151 |
+
"""
|
| 152 |
+
使用者回饋:將回饋資料寫入 Firestore
|
| 153 |
+
"""
|
| 154 |
try:
|
| 155 |
tz = pytz.timezone("Asia/Taipei")
|
| 156 |
timestamp_str = datetime.now(tz).strftime("%Y-%m-%d %H:%M:%S")
|
|
|
|
| 158 |
feedback["used_in_training"] = False
|
| 159 |
try:
|
| 160 |
db.collection("user_feedback").add(feedback)
|
| 161 |
+
except Exception:
|
| 162 |
pass
|
| 163 |
return {"message": "✅ 已記錄使用者回饋"}
|
| 164 |
except Exception as e:
|
|
|
|
| 168 |
# ──────────────────────────────────────────────────────────────────────────
|
| 169 |
# 強化 OCR 前處理 + 附帶 Debug 圖輸出
|
| 170 |
def preprocess_image_for_ocr(pil_image: Image.Image) -> Image.Image:
|
| 171 |
+
"""
|
| 172 |
+
前處理流程:
|
| 173 |
+
1. PIL Image → NumPy BGR
|
| 174 |
+
2. 灰階 + CLAHE(對比度增強)
|
| 175 |
+
3. 橘色背景遮罩 → 將背景橘色轉為白色
|
| 176 |
+
4. 固定閾值反向二值化
|
| 177 |
+
5. 放大 & GaussianBlur 平滑
|
| 178 |
+
中間各步驟會將影像存到 /tmp/debug_*.png,方便除錯
|
| 179 |
+
"""
|
| 180 |
+
# 1. PIL → NumPy (RGB->BGR)
|
| 181 |
img = np.array(pil_image.convert("RGB"))[:, :, ::-1]
|
| 182 |
|
| 183 |
+
# 2. 灰階 + CLAHE
|
| 184 |
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
|
| 185 |
clahe = cv2.createCLAHE(clipLimit=3.0, tileGridSize=(8, 8))
|
| 186 |
enhanced = clahe.apply(gray)
|
| 187 |
+
# Debug: CLAHE 增強後的灰階
|
|
|
|
| 188 |
Image.fromarray(enhanced).save("/tmp/debug_clahe.png")
|
| 189 |
|
| 190 |
+
# 3. HSV 色彩分離 (過濾橘色背景)
|
| 191 |
hsv = cv2.cvtColor(img, cv2.COLOR_BGR2HSV)
|
| 192 |
lower_orange = np.array([5, 100, 100])
|
| 193 |
upper_orange = np.array([20, 255, 255])
|
| 194 |
mask_orange = cv2.inRange(hsv, lower_orange, upper_orange)
|
| 195 |
+
# Debug: 橘色 mask
|
| 196 |
Image.fromarray(mask_orange).save("/tmp/debug_mask_orange.png")
|
| 197 |
|
| 198 |
+
# 將 mask 範圍內的像素設為白色(255),其餘保留灰階
|
| 199 |
filtered = enhanced.copy()
|
| 200 |
filtered[mask_orange > 0] = 255
|
| 201 |
+
# Debug: 過濾橘色後的灰階
|
|
|
|
| 202 |
Image.fromarray(filtered).save("/tmp/debug_filtered.png")
|
| 203 |
|
| 204 |
+
# 4. 固定閾值反向二值化 (threshold 200)
|
| 205 |
_, thresh = cv2.threshold(filtered, 200, 255, cv2.THRESH_BINARY_INV)
|
| 206 |
+
# Debug: 二值化後
|
| 207 |
Image.fromarray(thresh).save("/tmp/debug_thresh.png")
|
| 208 |
|
| 209 |
+
# 5. 放大 3 倍 & GaussianBlur 平滑
|
| 210 |
scaled = cv2.resize(thresh, None, fx=3.0, fy=3.0, interpolation=cv2.INTER_CUBIC)
|
| 211 |
smoothed = cv2.GaussianBlur(scaled, (5, 5), 0)
|
| 212 |
+
# Debug: 最終前處理結果
|
|
|
|
| 213 |
Image.fromarray(smoothed).save("/tmp/debug_processed.png")
|
| 214 |
|
| 215 |
return Image.fromarray(smoothed)
|
|
|
|
| 219 |
@app.post("/analyze-image")
|
| 220 |
async def analyze_uploaded_image(file: UploadFile = File(...)):
|
| 221 |
"""
|
| 222 |
+
圖片上傳並進行 OCR 辨識:
|
| 223 |
+
1. 讀取 UploadFile → PIL Image
|
| 224 |
+
2. 呼叫 preprocess_image_for_ocr 進行前處理 (並輸出 debug)
|
| 225 |
+
3. 用 pytesseract 擷取文字
|
| 226 |
+
4. 若擷取到文字,送給 BERT 做詐騙判斷
|
| 227 |
+
5. 回傳 JSON 包含 extracted_text 與 analysis_result
|
| 228 |
"""
|
| 229 |
+
# 1) 確認收到檔案
|
| 230 |
+
print("🔍 [DEBUG] 收到 analyze-image,檔名 =", file.filename)
|
| 231 |
+
|
| 232 |
try:
|
| 233 |
+
# 2) 讀取圖片 bytes,再轉成 PIL Image
|
| 234 |
image_bytes = await file.read()
|
| 235 |
+
print("🔍 [DEBUG] 圖片 bytes 長度 =", len(image_bytes))
|
| 236 |
+
pil_img = Image.open(io.BytesIO(image_bytes))
|
| 237 |
+
print("🔍 [DEBUG] PIL 成功開啟圖片,格式 =", pil_img.format, "大小 =", pil_img.size)
|
| 238 |
|
| 239 |
+
# 3) 強化前處理 (並產出 debug 影像)
|
| 240 |
+
processed_image = preprocess_image_for_ocr(pil_img)
|
| 241 |
|
| 242 |
+
# 4) Tesseract OCR
|
| 243 |
custom_config = r"-l chi_tra+eng --oem 3 --psm 6"
|
| 244 |
extracted_text = pytesseract.image_to_string(
|
| 245 |
processed_image,
|
| 246 |
config=custom_config
|
| 247 |
).strip()
|
| 248 |
+
print("🔍 [DEBUG] Tesseract 擷取文字 =", repr(extracted_text))
|
| 249 |
|
| 250 |
+
# 5) 如果沒有擷取到任何文字
|
| 251 |
if not extracted_text:
|
| 252 |
+
return JSONResponse({
|
| 253 |
"extracted_text": "",
|
| 254 |
"analysis_result": {
|
| 255 |
"status": "無法辨識",
|
| 256 |
"confidence": 0.0,
|
| 257 |
"suspicious_keywords": ["無法擷取分析結果"]
|
| 258 |
}
|
| 259 |
+
})
|
| 260 |
|
| 261 |
+
# 6) 擷取到文字後,呼叫 BERT 模型做詐騙判斷
|
| 262 |
result = bert_analyze_text(extracted_text)
|
| 263 |
+
return JSONResponse({
|
| 264 |
"extracted_text": extracted_text,
|
| 265 |
"analysis_result": result
|
| 266 |
+
})
|
| 267 |
+
|
| 268 |
except Exception as e:
|
| 269 |
+
# 印出詳細錯誤堆疊
|
| 270 |
+
import traceback
|
| 271 |
+
traceback.print_exc()
|
| 272 |
raise HTTPException(status_code=500, detail=f"圖片辨識失敗:{str(e)}")
|
| 273 |
|
| 274 |
|