Add Gaussian Noise Operation

Augmentor/Operations.py

@@ -1617,3 +1617,21 @@ def perform_operation(self, image):
          will expect an image of type PIL.Image)
         """
         return self.function_name(image, **self.function_arguments)
+
+class GaussianNoise(Operation):
+    """
+    The class `:class noise` is used to perfrom random noise on images passed
+    to its :func:`perform_operation` function.
+    """
+    def __init__(self, probability, mean, std):
+        Operation.__init__(self, probability)
+        self.mean = mean
+        self.std = std
+
+    def perform_operation(self, image):
+        w, h = image.size
+        c = len(image.getbands())
+
+        noise = np.random.normal(self.mean, self.std, (h, w, c))
+
+        return Image.fromarray(np.uint8(np.array(image) + noise))

Augmentor/Pipeline.py

@@ -1350,3 +1350,26 @@ def random_erasing(self, probability, rectangle_area):
             raise ValueError("The rectangle_area must be between 0.1 and 1.")
         else:
             self.add_operation(RandomErasing(probability=probability, rectangle_area=rectangle_area))
+
+    def noise(self, probability, mean=0, std=0.3):
+        """
+        The class `:class noise` is used to perfrom gaussian noise on images passed
+        to its :func:`perform_operation` function.
+
+        Its purpose is to make models robust to adversarial learning, by
+        randomly adding gassian noise to original images, which human could not
+        tell the difference between them.
+
+        :param probability: A value between 0 and 1 representing the
+         probability that the operation should be performed.
+        :param mean: The mean value of the gaussian distribution.
+        :param std: The std value of the gaussian distribution.
+        :return: None
+        """
+        if not 0 < probability <= 1:
+            raise ValueError(Pipeline._probability_error_text)
+        elif std <= 0:
+            raise ValueError('The standard deviation must larger than 0.')
+        else:
+            self.add_operation(GaussianNoise(probability=probability, mean=mean, std=std))
+

Augmentor/__init__.py

@@ -15,6 +15,6 @@
 
 __author__ = """Marcus D. Bloice"""
 __email__ = '[email protected]'
-__version__ = '0.1.10'
+__version__ = '0.1.11'
 
 __all__ = ['Pipeline']

README.md

@@ -208,6 +208,8 @@ p.flip_left_right(probability=0.8)
 p.flip_top_bottom(probability=0.3)
 p.crop_random(probability=1, percentage_area=0.5)
 p.resize(probability=1.0, width=120, height=120)
+p.noise(probability=1.0)
+# or p.noise(probability=1.0, mean=0, std=0.5)
 ```
 
 Once you have added the operations you require, you can sample images from this pipeline:

setup.py

@@ -3,7 +3,7 @@
 setup(
     name='Augmentor',
     packages=['Augmentor'],
-    version='0.1.10',
+    version='0.1.11',
     description='Image augmentation library for Machine Learning',
     long_description='Image augmentation library for Machine Learning',
     license='MIT',