Improve PIL and COLOR_BGR context image decode performance (#1501)

### Summary

- Ticket no. 141409
- There was a significant performance degradation on `decode_by_pil()`
if `ImageColorChannel.COLOR_BGR`.
- Improve it by returning a numpy array rather than `PIL.Image` and
using `cv2.cvtColor()` rather than `np.flip()`.

### How to test

- `tests/unit/test_image.py` covers this change.
- Compare performance with this code (dataset: Imagenette2)

| before | after |
| :-: | :-: |
|
![image](https://github.com/openvinotoolkit/datumaro/assets/26541465/3fd115b0-26b3-4a71-b058-0153cea96b98)
|
![image](https://github.com/openvinotoolkit/datumaro/assets/26541465/cb32adb6-cc70-4aef-9ab4-a47b862f1947)
|
| 00:51<00:00, 361.65it/s | 00:19<00:00, 859.39it/s |


### Checklist
<!-- Put an 'x' in all the boxes that apply -->
- [ ] I have added unit tests to cover my changes.​
- [ ] I have added integration tests to cover my changes.​
- [x] I have added the description of my changes into
[CHANGELOG](https://github.com/openvinotoolkit/datumaro/blob/develop/CHANGELOG.md).​
- [ ] I have updated the
[documentation](https://github.com/openvinotoolkit/datumaro/tree/develop/docs)
accordingly

### License

- [x] I submit _my code changes_ under the same [MIT
License](https://github.com/openvinotoolkit/datumaro/blob/develop/LICENSE)
that covers the project.
  Feel free to contact the maintainers if that's a concern.
- [x] I have updated the license header for each file (see an example
below).

```python
# Copyright (C) 2024 Intel Corporation
#
# SPDX-License-Identifier: MIT
```

---------

Signed-off-by: Kim, Vinnam <[email protected]>

CHANGELOG.md

@@ -25,6 +25,8 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   (<https://github.com/openvinotoolkit/datumaro/pull/1471>)
 - Add ExtractedMask and update importers who can use it to use it
   (<https://github.com/openvinotoolkit/datumaro/pull/1480>)
+- Improve PIL and COLOR_BGR context image decode performance
+  (<https://github.com/openvinotoolkit/datumaro/pull/1501>)
 
 ### Bug fixes
 - Split the video directory into subsets to avoid overwriting

src/datumaro/util/image.py

@@ -19,11 +19,12 @@
 from datumaro.components.crypter import NULL_CRYPTER, Crypter
 from datumaro.components.errors import DatumaroError
 
-try:
-    # Introduced in 1.20
-    from numpy.typing import DTypeLike
-except ImportError:
-    DTypeLike = Any
+if TYPE_CHECKING:
+    try:
+        # Introduced in 1.20
+        from numpy.typing import DTypeLike
+    except ImportError:
+        DTypeLike = Any
 
 
 class ImageBackend(Enum):
@@ -46,9 +47,6 @@ class ImageBackend(Enum):
 from datumaro.util.image_cache import ImageCache
 from datumaro.util.os_util import find_files
 
-if TYPE_CHECKING:
-    from PIL.Image import Image as PILImage
-
 
 class ImageColorChannel(Enum):
     """Image color channel
@@ -79,20 +77,21 @@ def decode_by_cv2(self, image_bytes: bytes) -> np.ndarray:
 
         raise ValueError
 
-    def decode_by_pil(self, image_bytes: bytes) -> PILImage:
+    def decode_by_pil(self, image_bytes: bytes) -> np.ndarray:
         """Convert image color channel for PIL Image."""
         from PIL import Image
 
         img = Image.open(BytesIO(image_bytes))
 
         if self == ImageColorChannel.UNCHANGED:
-            return img
+            return np.asarray(img)
 
         if self == ImageColorChannel.COLOR_BGR:
-            return Image.fromarray(np.flip(np.asarray(img.convert("RGB")), -1))
+            img = np.asarray(img.convert("RGB"))
+            return cv2.cvtColor(img, cv2.COLOR_RGB2BGR)
 
         if self == ImageColorChannel.COLOR_RGB:
-            return img.convert("RGB")
+            return np.asarray(img.convert("RGB"))
 
         raise ValueError
 
@@ -289,13 +288,13 @@ def decode_image(image_bytes: bytes, dtype: DTypeLike = np.uint8) -> np.ndarray:
 
     if IMAGE_BACKEND.get() == ImageBackend.cv2:
         image = ctx_color_scale.decode_by_cv2(image_bytes)
-        image = image.astype(dtype)
     elif IMAGE_BACKEND.get() == ImageBackend.PIL:
         image = ctx_color_scale.decode_by_pil(image_bytes)
-        image = np.asarray(image, dtype=dtype)
     else:
         raise NotImplementedError()
 
+    image = image.astype(dtype)
+
     assert len(image.shape) in {2, 3}
     if len(image.shape) == 3:
         assert image.shape[2] in {3, 4}