fix(watchlist): 采纳 #94 审查 — 防大图 OOM、OCR 不阻塞事件循环、收紧图片类型

为截图预处理增加像素上限与长边降采样；import-image 将 OCR 放入线程池；
去掉任意 image/* 放行，仅允许白名单 MIME/扩展名。

backend/app/api/watchlist.py

@@ -81,12 +81,15 @@ def ocr_status():
 @router.post("/import-image")
 async def import_from_image(request: Request, file: UploadFile = File(...)):
     """从自选截图识别股票代码，返回候选列表（不自动写入自选）。"""
+    import anyio
+
     content_type = (file.content_type or "").split(";")[0].strip().lower()
     filename = (file.filename or "").lower()
-    ok_type = content_type in _IMPORT_IMAGE_TYPES or content_type.startswith("image/")
+    # 严格白名单：不接受任意 image/*（如 image/svg+xml）
+    ok_type = content_type in _IMPORT_IMAGE_TYPES
     ok_ext = filename.endswith((".jpg", ".jpeg", ".png", ".webp", ".bmp", ".gif"))
     if not ok_type and not ok_ext:
-        raise HTTPException(400, "仅支持 JPG / PNG / WebP 等图片")
+        raise HTTPException(400, "仅支持 JPG / PNG / WebP / BMP / GIF 图片")
 
     data = await file.read()
     if not data:
@@ -97,7 +100,12 @@ async def import_from_image(request: Request, file: UploadFile = File(...)):
     existing = {r["symbol"] for r in watchlist.list_symbols()}
     data_dir = request.app.state.repo.store.data_dir
     try:
-        result = import_watchlist_image(data, data_dir, existing_symbols=existing)
+        # OCR 为同步 CPU/子进程；丢进线程池，避免卡住事件循环（行情 SSE 等）
+        result = await anyio.to_thread.run_sync(
+            lambda: import_watchlist_image(data, data_dir, existing_symbols=existing),
+        )
+    except ValueError as e:
+        raise HTTPException(400, str(e)) from e
     except RuntimeError as e:
         raise HTTPException(503, str(e)) from e
     except Exception as e:  # noqa: BLE001

backend/app/services/watchlist_ocr/provider.py

@@ -10,6 +10,13 @@ from PIL import Image, ImageEnhance, ImageOps
 
 logger = logging.getLogger(__name__)
 
+# 像素上限：防 JPEG 高分辨率解码后 OOM（字节上限挡不住高压缩大图）
+_MAX_PIXELS = 12_000_000
+# 长边上限：过大则降采样，减内存并加快 OCR
+_MAX_EDGE = 2000
+# 过窄时适度放大，提升小字识别率
+_MIN_WIDTH = 1400
+
 
 class OcrProvider(ABC):
     """OCR 引擎接口。实现类只负责「图 → 文本」，代码抽取与证券匹配在 pipeline 中完成。"""
@@ -26,19 +33,35 @@ class OcrProvider(ABC):
 
 
 def preprocess_for_ocr(image_bytes: bytes) -> Image.Image:
-    """暗色券商截图预处理：灰度、反相、增强对比，提升数字识别率。"""
+    """暗色券商截图预处理：像素上限、降采样、灰度、反相、增强对比。
+
+    Raises:
+        ValueError: 图片分辨率过高（像素数超限）。
+    """
     img = Image.open(BytesIO(image_bytes))
+    img.load()  # 强制解码，便于尽早失败 / 量尺寸
+
+    pixels = img.width * img.height
+    if pixels > _MAX_PIXELS:
+        raise ValueError("图片分辨率过高,请裁剪后重试")
+
+    # 大图降采样：既减内存又提速 OCR（只放大不缩小的旧逻辑已去掉）
+    if max(img.size) > _MAX_EDGE:
+        img.thumbnail((_MAX_EDGE, _MAX_EDGE), Image.Resampling.LANCZOS)
+
     if img.mode not in ("RGB", "L"):
         img = img.convert("RGB")
     gray = ImageOps.grayscale(img)
     # 暗底白字 → 白底黑字，Tesseract 更稳
     inverted = ImageOps.invert(gray)
     contrasted = ImageEnhance.Contrast(inverted).enhance(1.8)
-    # 小字放大一档
+    # 小字适度放大
     w, h = contrasted.size
-    if w < 1400:
-        scale = 1400 / w
-        contrasted = contrasted.resize((int(w * scale), int(h * scale)), Image.Resampling.LANCZOS)
+    if w < _MIN_WIDTH:
+        scale = _MIN_WIDTH / w
+        contrasted = contrasted.resize(
+            (int(w * scale), int(h * scale)), Image.Resampling.LANCZOS
+        )
     return contrasted
 
 

backend/tests/test_watchlist_ocr.py

@@ -1,16 +1,19 @@
 """自选截图 OCR：代码抽取与 instruments 匹配（不依赖本机 tesseract）。"""
 from __future__ import annotations
 
+from io import BytesIO
 from pathlib import Path
 
 import polars as pl
+import pytest
+from PIL import Image
 
 from app.services.watchlist_ocr.pipeline import (
     extract_codes,
     import_watchlist_image,
     resolve_candidates,
 )
-from app.services.watchlist_ocr.provider import OcrProvider
+from app.services.watchlist_ocr.provider import OcrProvider, preprocess_for_ocr
 
 
 class _FakeOcr(OcrProvider):
@@ -26,6 +29,12 @@ class _FakeOcr(OcrProvider):
         return self._text
 
 
+def _png_bytes(width: int, height: int) -> bytes:
+    buf = BytesIO()
+    Image.new("RGB", (width, height), color=(20, 20, 20)).save(buf, format="PNG")
+    return buf.getvalue()
+
+
 def test_extract_codes_order_and_dedupe():
     text = "563230 融\n515880\n价格 1.340\n563230 重复\nXAUUSD\n601636"
     assert extract_codes(text) == ["563230", "515880", "601636"]
@@ -73,3 +82,15 @@ def test_import_watchlist_image_with_fake_ocr(tmp_path: Path):
     assert result["codes"] == ["600036", "515880"]
     assert result["matched_count"] == 2
     assert result["unmatched_count"] == 0
+
+
+def test_preprocess_rejects_excessive_pixels():
+    # 4000×8000 = 32M 像素 > 12M 上限
+    with pytest.raises(ValueError, match="分辨率过高"):
+        preprocess_for_ocr(_png_bytes(4000, 8000))
+
+
+def test_preprocess_downsamples_large_edge():
+    # 2500×1000 = 2.5M 像素未超限，但长边 > 2000，应降采样
+    out = preprocess_for_ocr(_png_bytes(2500, 1000))
+    assert max(out.size) <= 2000