Face alignment with rotation using resample

[mowshon]

If you set the face alignment parameter to True, you can see pixel artefacts when you rotate the face horizontally over the eyes when working with faces of smaller dimensions.

Original Image

Face alignment

You can easily notice the changed pixels after rotation.

Rotate with resample=Image.BICUBIC

This PR adds a resample=Image.BICUBIC parameter to the align_face function, which smooths the changed pixels after face alignment.

What does it improve?

Firstly, there are now better facial images after calling DeepFace.extract_faces()

I initially had a theory that these deformed pixels are passed to the models to turn the face into vectors and thus affect the result for the worse.

I compared the two photos through all algorithms to see if smoothing via resample=Image.BICUBIC helps or not.

Image A	Image B

Code

result = DeepFace.verify(
    img1_path="musk-2.png",
    img2_path="musk-4.png",
    detector_backend=back,
    model_name=model,
    align=True
)

Each check was first run without Image.BICUBIC, then a second check with the same models was run with Image.BICUBIC

The entire report is at the link: https://docs.google.com/spreadsheets/d/1BPWT7bqz60QHu5y2Zx7Dhx0Kty4o_FDrVtf0vZYaTOw/edit?usp=sharing

• on the green background are the results where smoothing has improved the test result
• on the red background are the results where smoothing slightly worsened the result

Conclusions

This change affects faces from low resolution images. In large images - deformed pixels are not noticeable after rotation.

From the Excel report in the link above, it is noticeable that on the green background the percentage of difference is much higher than 12-19%. That is 12-19% improvement in the validation result between the two faces.

At the same time there is also a red background, but there the percentage of deterioration of the result is much lower.

[serengil]

I checked the impact of this change with current unit tests. Seems it is not changing the result positively or negatively. Accuracy score is exactly the same. This is because unit test items are high quality. So, if this does not have negative impact on current tests, good to go! Thank you for your contribution.

Before this change

https://github.com/serengil/deepface/actions/runs/9472574311/job/26101390979

Line 886, 24-06-11 22:28:47 - ✅ facial recognition models test passed with 99.33333333333333

After this change

https://github.com/serengil/deepface/actions/runs/9779715145/job/26999804690?pr=1269

Line 913, 24-07-03 14:39:09 - ✅ facial recognition models test passed with 99.33333333333333

[almirb]

I’m having trouble with face recognition on surveillance camera images (HD). Sometimes, the faces detected are distant from camera (low res faces)…. As I’m using align=True, maybe this change could help me.

Thanks @mowshon and @serengil