dev(narugo): add adversarial clean

docs/source/api_doc/restore/adversarial.rst

@@ -0,0 +1,14 @@
+imgutils.restore.adversarial
+====================================
+
+.. currentmodule:: imgutils.restore.adversarial
+
+.. automodule:: imgutils.restore.adversarial
+
+
+remove_adversarial_noise
+-----------------------------------
+
+.. autofunction:: remove_adversarial_noise
+
+

docs/source/api_doc/restore/adversarial_benchmark.plot.py

@@ -0,0 +1,28 @@
+import random
+
+from benchmark import BaseBenchmark, create_plot_cli
+from imgutils.restore.adversarial import remove_adversarial_noise
+
+
+class AdversarialRemovalBenchmark(BaseBenchmark):
+    def load(self):
+        pass
+
+    def unload(self):
+        pass
+
+    def run(self):
+        image_file = random.choice(self.all_images)
+
+        _ = remove_adversarial_noise(image_file)
+
+
+if __name__ == '__main__':
+    create_plot_cli(
+        [
+            ('Adversarial Removal', AdversarialRemovalBenchmark()),
+        ],
+        title='Benchmark for Adversarial Removal Algorithm',
+        run_times=5,
+        try_times=10,
+    )()

docs/source/api_doc/restore/adversarial_demo.plot.py

@@ -0,0 +1,18 @@
+import os
+
+from imgutils.data import load_image
+from imgutils.restore import remove_adversarial_noise
+from plot import image_plot
+
+sample_dir = 'sample'
+
+if __name__ == '__main__':
+    image = load_image(os.path.join(sample_dir, 'adversarial_input.png'))
+    image_plot(
+        [
+            (image, 'Adversarial Noised'),
+            (remove_adversarial_noise(image), 'Cleaned')
+        ],
+        columns=2,
+        figsize=(10, 6),
+    )

docs/source/api_doc/restore/index.rst

@@ -11,4 +11,5 @@ imgutils.restore
 
     nafnet
     scunet
+    adversarial
 

imgutils/restore/__init__.py

@@ -11,5 +11,6 @@
         :align: center
 
 """
+from .adversarial import remove_adversarial_noise
 from .nafnet import restore_with_nafnet
 from .scunet import restore_with_scunet

imgutils/restore/adversarial.py

@@ -0,0 +1,66 @@
+"""
+Overview:
+    Useful tools to remove adversarial noises, just using opencv library without any models.
+
+    .. image:: adversarial_demo.plot.py.svg
+        :align: center
+
+    This is an overall benchmark of all the adversarial denoising:
+
+    .. image:: adversarial_benchmark.plot.py.svg
+        :align: center
+
+    .. note::
+        This tool is inspired from `Github - lllyasviel/AdverseCleaner <https://github.com/lllyasviel/AdverseCleaner>`_.
+"""
+import cv2
+import numpy as np
+from PIL import Image
+from cv2.ximgproc import guidedFilter
+
+from ..data import load_image, ImageTyping
+
+
+def remove_adversarial_noise(image: ImageTyping, diameter: int = 5, sigma_color: float = 8.0,
+                             sigma_space: float = 8.0, radius: int = 4, eps: float = 16.0) -> Image.Image:
+    """
+    Remove adversarial noise from an image using bilateral and guided filtering.
+
+    This function applies bilateral filtering and guided filtering to reduce adversarial noise in the input image.
+
+    :param image: The input image.
+    :type image: ImageTyping
+
+    :param diameter: Diameter of each pixel neighborhood for bilateral filtering.
+    :type diameter: int, optional
+
+    :param sigma_color: Filter sigma in the color space for bilateral filtering.
+    :type sigma_color: float, optional
+
+    :param sigma_space: Filter sigma in the coordinate space for bilateral filtering.
+    :type sigma_space: float, optional
+
+    :param radius: Radius of Guided Filter.
+    :type radius: float, optional
+
+    :param eps: Guided Filter regularization term.
+    :type eps: int, optional
+
+    :return: Image with adversarial noise removed.
+    :rtype: Image.Image
+    """
+    image = load_image(image, mode='RGB', force_background='white')
+    img = np.array(image).astype(np.float32)
+    y = img.copy()
+
+    # Apply bilateral filtering iteratively
+    for _ in range(64):
+        y = cv2.bilateralFilter(y, diameter, sigma_color, sigma_space)
+
+    # Apply guided filtering iteratively
+    for _ in range(4):
+        y = guidedFilter(img, y, radius, eps)
+
+    # Clip the values and convert back to uint8 for PIL Image
+    output_image = Image.fromarray(y.clip(0, 255).astype(np.uint8))
+    return output_image

requirements.txt

@@ -5,6 +5,7 @@ scikit-learn
 huggingface_hub
 tqdm
 opencv-python
+opencv-contrib-python
 pandas
 scipy
 emoji>=2.5.0

test/restore/test_adversarial.py

@@ -0,0 +1,25 @@
+import pytest
+
+from imgutils.data import load_image
+from imgutils.restore import remove_adversarial_noise
+from test.testings import get_testfile
+
+
+@pytest.fixture()
+def adversarial_input():
+    return get_testfile('adversarial', 'adversarial_input.png')
+
+
+@pytest.fixture()
+def adversarial_output_pil():
+    return load_image(get_testfile('adversarial', 'adversarial_output.png'))
+
+
+@pytest.mark.unittest
+class TestRestoreAdversarial:
+    def test_remove_adversarial_noises(self, adversarial_input, adversarial_output_pil, image_diff):
+        assert image_diff(
+            remove_adversarial_noise(adversarial_input),
+            adversarial_output_pil,
+            throw_exception=False
+        ) < 1e-3