fix(python): honor EXIF orientation for bytes and PIL images (#755)

python/rapidocr/utils/load_image.py

@@ -81,14 +81,15 @@ class LoadImage:
             return img
 
         if isinstance(img, bytes):
-            img = self.img_to_ndarray(Image.open(BytesIO(img)))
+            img = self.exif_transpose(Image.open(BytesIO(img)))
+            img = self.img_to_ndarray(img)
             return img
 
         if isinstance(img, np.ndarray):
             return img
 
         if isinstance(img, Image.Image):
-            return self.img_to_ndarray(img)
+            return self.img_to_ndarray(self.exif_transpose(img))
 
         if isinstance(img, MemoryImage):
             try:
@@ -116,7 +117,7 @@ class LoadImage:
             if img_corrected is None:
                 return img
             return img_corrected
-        except Exception as e:
+        except Exception:
             return img
 
     def img_to_ndarray(self, img: Image.Image) -> np.ndarray:

python/tests/test_load_image_exif.py

@@ -0,0 +1,58 @@
+from pathlib import Path
+
+import numpy as np
+import pytest
+from PIL import Image
+from rapidocr.utils.load_image import LoadImage
+
+
+@pytest.mark.parametrize("input_type", ["bytes", "pil"])
+@pytest.mark.parametrize("orientation", [1, 3, 6, 8])
+def test_exif_input_types_match_path(tmp_path, input_type, orientation):
+    pixels = np.arange(18, dtype=np.uint8).reshape(2, 3, 3)
+    image = Image.fromarray(pixels)
+    exif = image.getexif()
+    exif[274] = orientation
+    path = tmp_path / "oriented.png"
+    image.save(path, exif=exif)
+    loader = LoadImage()
+    rotations = {1: 0, 3: 2, 6: 3, 8: 1}
+    expected = np.rot90(pixels, rotations[orientation])[..., ::-1]
+    np.testing.assert_array_equal(loader(path), expected)
+
+    if input_type == "bytes":
+        actual = loader(path.read_bytes())
+    else:
+        with Image.open(path) as pil_image:
+            original_pixels = np.array(pil_image)
+            actual = loader(pil_image)
+            assert pil_image.getexif()[274] == orientation
+            np.testing.assert_array_equal(np.array(pil_image), original_pixels)
+    np.testing.assert_array_equal(actual, expected)
+
+
+@pytest.mark.parametrize("input_type", ["bytes", "pil"])
+def test_exif_existing_jpeg_matches_path(input_type):
+    path = Path(__file__).parent / "test_files" / "img_exif_orientation.jpg"
+    loader = LoadImage()
+    expected = loader(path)
+    if input_type == "bytes":
+        actual = loader(path.read_bytes())
+    else:
+        with Image.open(path) as image:
+            actual = loader(image)
+    np.testing.assert_array_equal(actual, expected)
+
+
+@pytest.mark.parametrize("input_type", ["bytes", "pil"])
+def test_input_without_exif_preserves_pixels(tmp_path, input_type):
+    pixels = np.arange(18, dtype=np.uint8).reshape(2, 3, 3)
+    image = Image.fromarray(pixels)
+    path = tmp_path / "unoriented.png"
+    image.save(path)
+    loader = LoadImage()
+    if input_type == "bytes":
+        actual = loader(path.read_bytes())
+    else:
+        actual = loader(image)
+    np.testing.assert_array_equal(actual, pixels[..., ::-1])