!322 image transform method of real world robustness test

Merge pull request !322 from ZhidanLiu/master

examples/natural_robustness/natural_robustness_example.py

@@ -0,0 +1,176 @@
+# Copyright 2022 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Example for natural robustness methods."""
+
+import numpy as np
+import cv2
+
+from mindarmour.natural_robustness.image import Translate, Curve, Perspective, Scale, Shear, Rotate, SaltAndPepperNoise, \
+    NaturalNoise, GaussianNoise, UniformNoise, MotionBlur, GaussianBlur, GradientBlur, Contrast, GradientLuminance
+
+
+def test_perspective(image):
+    """Test perspective."""
+    ori_pos = [[0, 0], [0, 800], [800, 0], [800, 800]]
+    dst_pos = [[50, 0], [0, 800], [780, 0], [800, 800]]
+    trans = Perspective(ori_pos, dst_pos)
+    dst = trans(image)
+    cv2.imshow('dst', dst)
+    cv2.waitKey()
+
+
+def test_uniform_noise(image):
+    """Test uniform noise."""
+    trans = UniformNoise(factor=0.1)
+    dst = trans(image)
+    cv2.imshow('dst', dst)
+    cv2.waitKey()
+
+
+def test_gaussian_noise(image):
+    """Test gaussian noise."""
+    trans = GaussianNoise(factor=0.1)
+    dst = trans(image)
+    cv2.imshow('dst', dst)
+    cv2.waitKey()
+
+
+def test_contrast(image):
+    """Test contrast."""
+    trans = Contrast(alpha=2, beta=0)
+    dst = trans(image)
+    cv2.imshow('dst', dst)
+    cv2.waitKey()
+
+
+def test_gaussian_blur(image):
+    """Test gaussian blur."""
+    trans = GaussianBlur(ksize=5)
+    dst = trans(image)
+    cv2.imshow('dst', dst)
+    cv2.waitKey()
+
+
+def test_salt_and_pepper_noise(image):
+    """Test salt and pepper noise."""
+    trans = SaltAndPepperNoise(factor=0.01)
+    dst = trans(image)
+    cv2.imshow('dst', dst)
+    cv2.waitKey()
+
+
+def test_translate(image):
+    """Test translate."""
+    trans = Translate(x_bias=0.1, y_bias=0.1)
+    dst = trans(image)
+    cv2.imshow('dst', dst)
+    cv2.waitKey()
+
+
+def test_scale(image):
+    """Test scale."""
+    trans = Scale(factor_x=0.7, factor_y=0.7)
+    dst = trans(image)
+    cv2.imshow('dst', dst)
+    cv2.waitKey()
+
+
+def test_shear(image):
+    """Test shear."""
+    trans = Shear(factor=0.2)
+    dst = trans(image)
+    cv2.imshow('dst', dst)
+    cv2.waitKey()
+
+
+def test_rotate(image):
+    """Test rotate."""
+    trans = Rotate(angle=20)
+    dst = trans(image)
+    cv2.imshow('dst', dst)
+    cv2.waitKey()
+
+
+def test_curve(image):
+    """Test curve."""
+    trans = Curve(curves=2, depth=1.5, mode='horizontal')
+    dst = trans(image)
+    cv2.imshow('dst', dst)
+    cv2.waitKey()
+
+
+def test_natural_noise(image):
+    """Test natural noise."""
+    trans = NaturalNoise(ratio=0.0001, k_x_range=(1, 30), k_y_range=(1, 10), auto_param=True)
+    dst = trans(image)
+    cv2.imshow('dst', dst)
+    cv2.waitKey()
+
+
+def test_gradient_luminance(image):
+    """Test gradient luminance."""
+    height, width = image.shape[:2]
+    point = (height // 4, width // 2)
+    start = (255, 255, 255)
+    end = (0, 0, 0)
+    scope = 0.3
+    bright_rate = 0.4
+    trans = GradientLuminance(start, end, start_point=point, scope=scope, pattern='dark', bright_rate=bright_rate,
+                              mode='horizontal')
+    dst = trans(image)
+    cv2.imshow('dst', dst)
+    cv2.waitKey()
+
+
+def test_motion_blur(image):
+    """Test motion blur."""
+    angle = -10.5
+    i = 10
+    trans = MotionBlur(degree=i, angle=angle)
+    dst = trans(image)
+    cv2.imshow('dst', dst)
+    cv2.waitKey()
+
+
+def test_gradient_blur(image):
+    """Test gradient blur."""
+    number = 10
+    h, w = image.shape[:2]
+    point = (int(h / 5), int(w / 5))
+    center = False
+    trans = GradientBlur(point, number, center)
+    dst = trans(image)
+    cv2.imshow('dst', dst)
+    cv2.waitKey()
+
+
+if __name__ == '__main__':
+    img = cv2.imread('1.jpeg')
+    img = np.array(img)
+    test_uniform_noise(img)
+    test_gaussian_noise(img)
+    test_motion_blur(img)
+    test_gradient_blur(img)
+    test_gradient_luminance(img)      #####
+    test_natural_noise(img)
+    test_curve(img)
+    test_rotate(img)
+    test_shear(img)
+    test_scale(img)
+    test_translate(img)
+    test_salt_and_pepper_noise(img)
+    test_gaussian_blur(img)
+    test_contrast(img)
+    test_perspective(img)

mindarmour/natural_robustness/__init__.py

@@ -0,0 +1,18 @@
+# Copyright 2022 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Natural Robustness.
+
+This module is a tool box to enhance model security and against natural perturb examples.
+"""

mindarmour/natural_robustness/image/__init__.py

@@ -0,0 +1,37 @@
+# Copyright 2022 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+This package include methods to generate natural perturbation samples.
+"""
+
+from .transformation import Translate, Scale, Shear, Rotate, Perspective, Curve
+from .blur import GaussianBlur, MotionBlur, GradientBlur
+from .luminance import Contrast, GradientLuminance
+from .corruption import UniformNoise, GaussianNoise, SaltAndPepperNoise, NaturalNoise
+
+__all__ = ['Translate',
+           'Scale',
+           'Shear',
+           'Rotate',
+           'Perspective',
+           'Curve',
+           'GaussianBlur',
+           'MotionBlur',
+           'GradientBlur',
+           'Contrast',
+           'GradientLuminance',
+           'UniformNoise',
+           'GaussianNoise',
+           'SaltAndPepperNoise',
+           'NaturalNoise']

mindarmour/natural_robustness/image/blur.py

@@ -0,0 +1,193 @@
+# Copyright 2022 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Image Blur
+"""
+
+import numpy as np
+import cv2
+
+from mindarmour.natural_robustness.image.natural_perturb import _NaturalPerturb
+from mindarmour.utils._check_param import check_param_multi_types, check_int_positive, check_param_type
+from mindarmour.utils.logger import LogUtil
+
+LOGGER = LogUtil.get_instance()
+TAG = 'Image Blur'
+
+
+class GaussianBlur(_NaturalPerturb):
+    """
+    Blurs the image using Gaussian blur filter.
+
+    Args:
+        ksize (int): Size of gaussian kernel, this value must be non-negnative.
+        auto_param (bool): Auto selected parameters. Selected parameters will preserve semantics of image.
+
+    Example:
+        >>> img = cv2.imread('1.png')
+        >>> img = np.array(img)
+        >>> ksize = 5
+        >>> trans = GaussianBlur(ksize)
+        >>> dst = trans(img)
+    """
+
+    def __init__(self, ksize=2, auto_param=False):
+        super(GaussianBlur, self).__init__()
+        ksize = check_int_positive('ksize', ksize)
+        if auto_param:
+            ksize = 2 * np.random.randint(0, 5) + 1
+        else:
+            ksize = 2 * ksize + 1
+        self.ksize = (ksize, ksize)
+
+    def __call__(self, image):
+        """
+        Transform the image.
+
+        Args:
+            image (numpy.ndarray): Original image to be transformed.
+
+        Returns:
+            numpy.ndarray, transformed image.
+        """
+        ori_dtype = image.dtype
+        _, chw, normalized, gray3dim, image = self._check(image)
+        new_img = cv2.GaussianBlur(image, self.ksize, 0)
+        new_img = self._original_format(new_img, chw, normalized, gray3dim)
+        return new_img.astype(ori_dtype)
+
+
+class MotionBlur(_NaturalPerturb):
+    """
+    Motion blur for a given image.
+
+    Args:
+        degree (int): Degree of blur. This value must be positive. Suggested value range in [1, 15].
+        angle: (union[float, int]): Direction of motion blur. Angle=0 means up and down motion blur. Angle is
+            counterclockwise.
+        auto_param (bool): Auto selected parameters. Selected parameters will preserve semantics of image.
+
+    Example:
+        >>> img = cv2.imread('1.png')
+        >>> img = np.array(img)
+        >>> angle = 0
+        >>> degree = 5
+        >>> trans = MotionBlur(degree=degree, angle=angle)
+        >>> new_img = trans(img)
+    """
+
+    def __init__(self, degree=5, angle=45, auto_param=False):
+        super(MotionBlur, self).__init__()
+        self.degree = check_int_positive('degree', degree)
+        self.degree = check_param_multi_types('degree', degree, [float, int])
+        auto_param = check_param_type('auto_param', auto_param, bool)
+        if auto_param:
+            self.degree = np.random.randint(1, 5)
+            self.angle = np.random.uniform(0, 360)
+        else:
+            self.angle = angle - 45
+
+    def __call__(self, image):
+        """
+        Motion blur for a given image.
+
+        Args:
+            image (numpy.ndarray): Original image.
+
+        Returns:
+            numpy.ndarray, image after motion blur.
+        """
+        ori_dtype = image.dtype
+        _, chw, normalized, gray3dim, image = self._check(image)
+        matrix = cv2.getRotationMatrix2D((self.degree / 2, self.degree / 2), self.angle, 1)
+        motion_blur_kernel = np.diag(np.ones(self.degree))
+        motion_blur_kernel = cv2.warpAffine(motion_blur_kernel, matrix, (self.degree, self.degree))
+        motion_blur_kernel = motion_blur_kernel / self.degree
+        blurred = cv2.filter2D(image, -1, motion_blur_kernel)
+        # convert to uint8
+        cv2.normalize(blurred, blurred, 0, 255, cv2.NORM_MINMAX)
+        blurred = self._original_format(blurred, chw, normalized, gray3dim)
+
+        return blurred.astype(ori_dtype)
+
+
+class GradientBlur(_NaturalPerturb):
+    """
+    Gradient blur.
+
+    Args:
+        point (union[tuple, list]): 2D coordinate of the Blur center point.
+        kernel_num (int): Number of blur kernels. Suggested value range in [1, 8].
+        center (bool): Blurred or clear at the center of a specified point.
+        auto_param (bool): Auto selected parameters. Selected parameters will preserve semantics of image.
+
+    Example:
+        >>> img = cv2.imread('xx.png')
+        >>> img = np.array(img)
+        >>> number = 5
+        >>> h, w = img.shape[:2]
+        >>> point = (int(h / 5), int(w / 5))
+        >>> center = True
+        >>> trans = GradientBlur(point, number,  center)
+        >>> new_img = trans(img)
+    """
+
+    def __init__(self, point, kernel_num=3, center=True, auto_param=False):
+        super(GradientBlur).__init__()
+        point = check_param_multi_types('point', point, [list, tuple])
+        self.auto_param = check_param_type('auto_param', auto_param, bool)
+        self.point = tuple(point)
+        self.kernel_num = check_int_positive('kernel_num', kernel_num)
+        self.center = check_param_type('center', center, bool)
+
+    def _auto_param(self, h, w):
+        self.point = (int(np.random.uniform(0, h)), int(np.random.uniform(0, w)))
+        self.kernel_num = np.random.randint(1, 6)
+        self.center = np.random.choice([True, False])
+
+    def __call__(self, image):
+        """
+        Args:
+            image (numpy.ndarray): Original image.
+
+        Returns:
+            numpy.ndarray, gradient blurred image.
+        """
+        ori_dtype = image.dtype
+        _, chw, normalized, gray3dim, image = self._check(image)
+        w, h = image.shape[:2]
+        if self.auto_param:
+            self._auto_param(h, w)
+        mask = np.zeros(image.shape, dtype=np.uint8)
+        masks = []
+        radius = max(w - self.point[0], self.point[0], h - self.point[1], self.point[1])
+        radius = int(radius / self.kernel_num)
+        for i in range(self.kernel_num):
+            circle = cv2.circle(mask.copy(), self.point, radius * (1 + i), (1, 1, 1), -1)
+            masks.append(circle)
+        blurs = []
+        for i in range(3, 3 + 2 * self.kernel_num, 2):
+            ksize = (i, i)
+            blur = cv2.GaussianBlur(image, ksize, 0)
+            blurs.append(blur)
+
+        dst = image.copy()
+        if self.center:
+            for i in range(self.kernel_num):
+                dst = masks[i] * dst + (1 - masks[i]) * blurs[i]
+        else:
+            for i in range(self.kernel_num - 1, -1, -1):
+                dst = masks[i] * blurs[self.kernel_num - 1 - i] + (1 - masks[i]) * dst
+        dst = self._original_format(dst, chw, normalized, gray3dim)
+        return dst.astype(ori_dtype)

mindarmour/natural_robustness/image/corruption.py

@@ -0,0 +1,251 @@
+# Copyright 2022 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Image corruption.
+"""
+import math
+import numpy as np
+import cv2
+
+from mindarmour.natural_robustness.image.natural_perturb import _NaturalPerturb
+from mindarmour.utils._check_param import check_param_multi_types, check_param_type
+from mindarmour.utils.logger import LogUtil
+
+LOGGER = LogUtil.get_instance()
+TAG = 'Image corruption'
+
+
+class UniformNoise(_NaturalPerturb):
+    """
+    Add uniform noise of an image.
+
+    Args:
+        factor (float): Noise density, the proportion of noise points per unit pixel area. Suggested value range in
+            [0.001, 0.15].
+        auto_param (bool): Auto selected parameters. Selected parameters will preserve semantics of image.
+
+    Example:
+        >>> img = cv2.imread('1.png')
+        >>> img = np.array(img)
+        >>> factor = 0.1
+        >>> trans = UniformNoise(factor)
+        >>> dst = trans(img)
+    """
+
+    def __init__(self, factor=0.1, auto_param=False):
+        super(UniformNoise, self).__init__()
+        self.factor = check_param_multi_types('factor', factor, [int, float])
+        check_param_type('auto_param', auto_param, bool)
+        if auto_param:
+            self.factor = np.random.uniform(0, 0.15)
+
+    def __call__(self, image):
+        """
+        Transform the image.
+
+        Args:
+            image (numpy.ndarray): Original image to be transformed.
+
+        Returns:
+            numpy.ndarray, transformed image.
+        """
+        ori_dtype = image.dtype
+        _, chw, normalized, gray3dim, image = self._check(image)
+        low, high = (0, 255)
+        weight = self.factor * (high - low)
+        noise = np.random.uniform(-weight, weight, size=image.shape)
+        trans_image = np.clip(image + noise, low, high)
+        trans_image = self._original_format(trans_image, chw, normalized, gray3dim)
+
+        return trans_image.astype(ori_dtype)
+
+
+class GaussianNoise(_NaturalPerturb):
+    """
+    Add gaussian noise of an image.
+
+    Args:
+        factor (float): Noise density, the proportion of noise points per unit pixel area. Suggested value range in
+            [0.001, 0.15].
+        auto_param (bool): Auto selected parameters. Selected parameters will preserve semantics of image.
+
+    Example:
+        >>> img = cv2.imread('1.png')
+        >>> img = np.array(img)
+        >>> factor = 0.1
+        >>> trans = GaussianNoise(factor)
+        >>> dst = trans(img)
+    """
+
+    def __init__(self, factor=0.1, auto_param=False):
+        super(GaussianNoise, self).__init__()
+        self.factor = check_param_multi_types('factor', factor, [int, float])
+        check_param_type('auto_param', auto_param, bool)
+        if auto_param:
+            self.factor = np.random.uniform(0, 0.15)
+
+    def __call__(self, image):
+        """
+        Transform the image.
+
+        Args:
+            image (numpy.ndarray): Original image to be transformed.
+
+        Returns:
+            numpy.ndarray, transformed image.
+        """
+        ori_dtype = image.dtype
+        low, high = (0, 255)
+        _, chw, normalized, gray3dim, image = self._check(image)
+        std = self.factor / math.sqrt(3) * (high - low)
+        noise = np.random.normal(scale=std, size=image.shape)
+        trans_image = np.clip(image + noise, low, high)
+        trans_image = self._original_format(trans_image, chw, normalized, gray3dim)
+        return trans_image.astype(ori_dtype)
+
+
+class SaltAndPepperNoise(_NaturalPerturb):
+    """
+    Add salt and pepper noise of an image.
+
+    Args:
+        factor (float): Noise density, the proportion of noise points per unit pixel area. Suggested value range in
+            [0.001, 0.15].
+        auto_param (bool): Auto selected parameters. Selected parameters will preserve semantics of image.
+
+    Example:
+        >>> img = cv2.imread('1.png')
+        >>> img = np.array(img)
+        >>> factor = 0.1
+        >>> trans = SaltAndPepperNoise(factor)
+        >>> dst = trans(img)
+    """
+
+    def __init__(self, factor=0, auto_param=False):
+        super(SaltAndPepperNoise, self).__init__()
+        self.factor = check_param_multi_types('factor', factor, [int, float])
+        check_param_type('auto_param', auto_param, bool)
+        if auto_param:
+            self.factor = np.random.uniform(0, 0.15)
+
+    def __call__(self, image):
+        """
+        Transform the image.
+
+        Args:
+            image (numpy.ndarray): Original image to be transformed.
+
+        Returns:
+            numpy.ndarray, transformed image.
+        """
+        ori_dtype = image.dtype
+        _, chw, normalized, gray3dim, image = self._check(image)
+        low, high = (0, 255)
+        noise = np.random.uniform(low=-1, high=1, size=(image.shape[0], image.shape[1]))
+        trans_image = np.copy(image)
+        threshold = 1 - self.factor
+        trans_image[noise < -threshold] = low
+        trans_image[noise > threshold] = high
+        trans_image = self._original_format(trans_image, chw, normalized, gray3dim)
+        return trans_image.astype(ori_dtype)
+
+
+class NaturalNoise(_NaturalPerturb):
+    """
+    Add natural noise to an image.
+
+    Args:
+        ratio (float): Noise density, the proportion of noise blocks per unit pixel area. Suggested value range in
+            [0.00001, 0.001].
+        k_x_range (union[list, tuple]): Value range of the noise block length.
+        k_y_range (union[list, tuple]): Value range of the noise block width.
+        auto_param (bool): Auto selected parameters. Selected parameters will preserve semantics of image.
+
+    Examples:
+        >>> img = cv2.imread('xx.png')
+        >>> img = np.array(img)
+        >>> ratio = 0.0002
+        >>> k_x_range = (1, 5)
+        >>> k_y_range = (3, 25)
+        >>> trans = NaturalNoise(ratio, k_x_range, k_y_range)
+        >>> new_img = trans(img)
+    """
+
+    def __init__(self, ratio=0.0002, k_x_range=(1, 5), k_y_range=(3, 25), auto_param=False):
+        super(NaturalNoise).__init__()
+        self.ratio = check_param_type('ratio', ratio, float)
+        k_x_range = check_param_multi_types('k_x_range', k_x_range, [list, tuple])
+        k_y_range = check_param_multi_types('k_y_range', k_y_range, [list, tuple])
+        self.k_x_range = tuple(k_x_range)
+        self.k_y_range = tuple(k_y_range)
+        self.auto_param = check_param_type('auto_param', auto_param, bool)
+
+    def __call__(self, image):
+        """
+        Add natural noise to given image.
+
+        Args:
+            image (numpy.ndarray): Original image.
+
+        Returns:
+            numpy.ndarray, image with natural noise.
+        """
+        ori_dtype = image.dtype
+        _, chw, normalized, gray3dim, image = self._check(image)
+        randon_range = 100
+        w, h = image.shape[:2]
+        channel = len(np.shape(image))
+
+        if self.auto_param:
+            self.ratio = np.random.uniform(0, 0.001)
+            self.k_x_range = (1, 0.1 * w)
+            self.k_y_range = (1, 0.1 * h)
+
+        for _ in range(5):
+            if channel == 3:
+                noise = np.ones((w, h, 3), dtype=np.uint8) * 255
+                dst = np.ones((w, h, 3), dtype=np.uint8) * 255
+            else:
+                noise = np.ones((w, h), dtype=np.uint8) * 255
+                dst = np.ones((w, h), dtype=np.uint8) * 255
+
+            rate = self.ratio / 5
+            mask = np.random.uniform(size=(w, h)) < rate
+            noise[mask] = np.random.randint(0, randon_range)
+
+            k_x, k_y = np.random.randint(*self.k_x_range), np.random.randint(*self.k_y_range)
+            kernel = np.ones((k_x, k_y), np.uint8)
+            erode = cv2.erode(noise, kernel, iterations=1)
+            dst = erode * (erode < randon_range) + dst * (1 - erode < randon_range)
+            # Add black point
+            for _ in range(np.random.randint(math.ceil(k_x * k_y / 2))):
+                x = np.random.randint(-k_x, k_x)
+                y = np.random.randint(-k_y, k_y)
+                matrix = np.array([[1, 0, y], [0, 1, x]], dtype=np.float)
+                affine = cv2.warpAffine(noise, matrix, (h, w))
+                dst = affine * (affine < randon_range) + dst * (1 - affine < randon_range)
+            # Add white point
+            for _ in range(int(k_x * k_y / 2)):
+                x = np.random.randint(-k_x / 2 - 1, k_x / 2 + 1)
+                y = np.random.randint(-k_y / 2 - 1, k_y / 2 + 1)
+                matrix = np.array([[1, 0, y], [0, 1, x]], dtype=np.float)
+                affine = cv2.warpAffine(noise, matrix, (h, w))
+                white = affine < randon_range
+                dst[white] = 255
+
+        mask = dst < randon_range
+        dst = image * (1 - mask) + dst * mask
+        dst = self._original_format(dst, chw, normalized, gray3dim)
+
+        return dst.astype(ori_dtype)

mindarmour/natural_robustness/image/luminance.py

@@ -0,0 +1,287 @@
+# Copyright 2022 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Image luminance.
+"""
+import math
+import numpy as np
+import cv2
+
+from mindarmour.natural_robustness.image.natural_perturb import _NaturalPerturb
+from mindarmour.utils._check_param import check_param_multi_types, check_param_in_range, check_param_type, \
+    check_value_non_negative
+from mindarmour.utils.logger import LogUtil
+
+LOGGER = LogUtil.get_instance()
+TAG = 'Image Luminance'
+
+
+class Contrast(_NaturalPerturb):
+    """
+    Contrast of an image.
+
+    Args:
+        alpha (Union[float, int]): Control the contrast of an image. :math:`out_image = in_image*alpha+beta`.
+            Suggested value range in [0.2, 2].
+        beta (Union[float, int]): Delta added to alpha. Default: 0.
+        auto_param (bool): Auto selected parameters. Selected parameters will preserve semantics of image.
+
+    Example:
+        >>> img = cv2.imread('1.png')
+        >>> img = np.array(img)
+        >>> alpha = 0.1
+        >>> beta = 1
+        >>> trans = Contrast(alpha, beta)
+        >>> dst = trans(img)
+    """
+
+    def __init__(self, alpha=1, beta=0, auto_param=False):
+        super(Contrast, self).__init__()
+        self.alpha = check_param_multi_types('factor', alpha, [int, float])
+        self.beta = check_param_multi_types('factor', beta, [int, float])
+        auto_param = check_param_type('auto_param', auto_param, bool)
+        if auto_param:
+            self.alpha = np.random.uniform(0.2, 2)
+            self.beta = np.random.uniform(-20, 20)
+
+    def __call__(self, image):
+        """
+        Transform the image.
+
+        Args:
+            image (numpy.ndarray): Original image to be transformed.
+
+        Returns:
+            numpy.ndarray, transformed image.
+        """
+        ori_dtype = image.dtype
+        _, chw, normalized, gray3dim, image = self._check(image)
+        dst = cv2.convertScaleAbs(image, alpha=self.alpha, beta=self.beta)
+        dst = self._original_format(dst, chw, normalized, gray3dim)
+        return dst.astype(ori_dtype)
+
+
+def _circle_gradient_mask(img_src, color_start, color_end, scope=0.5, point=None):
+    """
+    Generate circle gradient mask.
+
+    Args:
+        img_src (numpy.ndarray): Source image.
+        color_start (union([tuple, list])): Color of circle gradient center.
+        color_end (union([tuple, list])): Color of circle gradient edge.
+        scope (float): Range of the gradient. A larger value indicates a larger gradient range.
+        point (union([tuple, list]): Gradient center point.
+
+    Returns:
+        numpy.ndarray, gradients mask.
+    """
+    if not isinstance(img_src, np.ndarray):
+        raise TypeError('`src` must be numpy.ndarray type, but got {0}.'.format(type(img_src)))
+
+    shape = img_src.shape
+    height, width = shape[:2]
+    rgb = False
+    if len(shape) == 3:
+        rgb = True
+    if point is None:
+        point = (height // 2, width // 2)
+    x, y = point
+
+    # upper left
+    bound_upper_left = math.ceil(math.sqrt(x ** 2 + y ** 2))
+    # upper right
+    bound_upper_right = math.ceil(math.sqrt(height ** 2 + (width - y) ** 2))
+    # lower left
+    bound_lower_left = math.ceil(math.sqrt((height - x) ** 2 + y ** 2))
+    # lower right
+    bound_lower_right = math.ceil(math.sqrt((height - x) ** 2 + (width - y) ** 2))
+
+    radius = max(bound_lower_left, bound_lower_right, bound_upper_left, bound_upper_right) * scope
+
+    img_grad = np.ones_like(img_src, dtype=np.uint8) * max(color_end)
+    # opencv use BGR format
+    grad_b = float(color_end[0] - color_start[0]) / radius
+    grad_g = float(color_end[1] - color_start[1]) / radius
+    grad_r = float(color_end[2] - color_start[2]) / radius
+
+    for i in range(height):
+        for j in range(width):
+            distance = math.ceil(math.sqrt((x - i) ** 2 + (y - j) ** 2))
+            if distance >= radius:
+                continue
+            if rgb:
+                img_grad[i, j, 0] = color_start[0] + distance * grad_b
+                img_grad[i, j, 1] = color_start[1] + distance * grad_g
+                img_grad[i, j, 2] = color_start[2] + distance * grad_r
+            else:
+                img_grad[i, j] = color_start[0] + distance * grad_b
+
+    return img_grad.astype(np.uint8)
+
+
+def _line_gradient_mask(image, start_pos=None, start_color=(0, 0, 0), end_color=(255, 255, 255), mode='horizontal'):
+    """
+    Generate liner gradient mask.
+
+    Args:
+        image (numpy.ndarray): Original image.
+        start_pos (union[tuple, list]): 2D coordinate of gradient center.
+        start_color (union([tuple, list])): Color of circle gradient center.
+        end_color (union([tuple, list])): Color of circle gradient edge.
+        mode　(str): Direction of gradient. Optional value is 'vertical' or 'horizontal'.
+
+    Returns:
+        numpy.ndarray, gradients mask.
+    """
+    shape = image.shape
+    h, w = shape[:2]
+    rgb = False
+    if len(shape) == 3:
+        rgb = True
+    if start_pos is None:
+        start_pos = 0.5
+    else:
+        if mode == 'horizontal':
+            if start_pos[0] > h:
+                start_pos = 1
+            else:
+                start_pos = start_pos[0] / h
+        else:
+            if start_pos[1] > w:
+                start_pos = 1
+            else:
+                start_pos = start_pos[1] / w
+    start_color = np.array(start_color)
+    end_color = np.array(end_color)
+    if mode == 'horizontal':
+        w_l = int(w * start_pos)
+        w_r = w - w_l
+        if w_l > w_r:
+            r_end_color = (end_color - start_color) / start_pos * (1 - start_pos) + start_color
+            left = np.linspace(end_color, start_color, w_l)
+            right = np.linspace(start_color, r_end_color, w_r)
+        else:
+            l_end_color = (end_color - start_color) / (1 - start_pos) * start_pos + start_color
+            left = np.linspace(l_end_color, start_color, w_l)
+            right = np.linspace(start_color, end_color, w_r)
+        line = np.concatenate((left, right), axis=0)
+        mask = np.reshape(np.tile(line, (h, 1)), (h, w, 3))
+    else:
+        # 'vertical'
+        h_t = int(h * start_pos)
+        h_b = h - h_t
+        if h_t > h_b:
+            b_end_color = (end_color - start_color) / start_pos * (1 - start_pos) + start_color
+            top = np.linspace(end_color, start_color, h_t)
+            bottom = np.linspace(start_color, b_end_color, h_b)
+        else:
+            t_end_color = (end_color - start_color) / (1 - start_pos) * start_pos + start_color
+            top = np.linspace(t_end_color, start_color, h_t)
+            bottom = np.linspace(start_color, end_color, h_b)
+        line = np.concatenate((top, bottom), axis=0)
+        mask = np.reshape(np.tile(line, (w, 1)), (w, h, 3))
+        mask = np.transpose(mask, [1, 0, 2])
+    if not rgb:
+        mask = mask[:, :, 0]
+    return mask.astype(np.uint8)
+
+
+class GradientLuminance(_NaturalPerturb):
+    """
+    Gradient adjusts the luminance of picture.
+
+    Args:
+        color_start (union[tuple, list]): Color of gradient center. Default:(0, 0, 0).
+        color_end (union[tuple, list]):  Color of gradient edge. Default:(255, 255, 255).
+        start_point (union[tuple, list]): 2D coordinate of gradient center.
+        scope (float): Range of the gradient. A larger value indicates a larger gradient range. Default: 0.3.
+        pattern (str): Dark or light, this value must be in ['light', 'dark'].
+        bright_rate (float): Control brightness. A larger value indicates a larger gradient range. If parameter
+            'pattern' is 'light', Suggested value range in [0.1, 0.7], if parameter 'pattern' is 'dark', Suggested value
+            range in [0.1, 0.9].
+        mode (str): Gradient mode, value must be in ['circle', 'horizontal', 'vertical'].
+        auto_param (bool): Auto selected parameters. Selected parameters will preserve semantics of image.
+
+    Examples:
+        >>> img = cv2.imread('x.png')
+        >>> height, width = img.shape[:2]
+        >>> point = (height // 4, width // 2)
+        >>> start = (255, 255, 255)
+        >>> end = (0, 0, 0)
+        >>> scope = 0.3
+        >>> pattern='light'
+        >>> bright_rate = 0.3
+        >>> trans = GradientLuminance(start, end, point, scope, pattern, bright_rate, mode='circle')
+        >>> img_new = trans(img)
+    """
+
+    def __init__(self, color_start=(0, 0, 0), color_end=(255, 255, 255), start_point=(10, 10), scope=0.5,
+                 pattern='light', bright_rate=0.3, mode='circle', auto_param=False):
+        super(GradientLuminance, self).__init__()
+        self.color_start = check_param_multi_types('color_start', color_start, [list, tuple])
+        self.color_end = check_param_multi_types('color_end', color_end, [list, tuple])
+        self.start_point = check_param_multi_types('start_point', start_point, [list, tuple])
+        self.scope = check_value_non_negative('scope', scope)
+        self.bright_rate = check_param_type('bright_rate', bright_rate, float)
+        self.bright_rate = check_param_in_range('bright_rate', bright_rate, 0, 1)
+        self.auto_param = check_param_type('auto_param', auto_param, bool)
+
+        if pattern in ['light', 'dark']:
+            self.pattern = pattern
+        else:
+            msg = "Value of param pattern must be in ['light', 'dark']"
+            LOGGER.error(TAG, msg)
+            raise ValueError(msg)
+        if mode in ['circle', 'horizontal', 'vertical']:
+            self.mode = mode
+        else:
+            msg = "Value of param mode must be in ['circle', 'horizontal', 'vertical']"
+            LOGGER.error(TAG, msg)
+            raise ValueError(msg)
+
+    def _set_auto_param(self, w, h):
+        self.color_start = (np.random.uniform(0, 255),) * 3
+        self.color_end = (np.random.uniform(0, 255),) * 3
+        self.start_point = (np.random.uniform(0, w), np.random.uniform(0, h))
+        self.scope = np.random.uniform(0, 1)
+        self.bright_rate = np.random.uniform(0.1, 0.9)
+        self.pattern = np.random.choice(['light', 'dark'])
+        self.mode = np.random.choice(['circle', 'horizontal', 'vertical'])
+
+    def __call__(self, image):
+        """
+        Gradient adjusts the luminance of picture.
+
+        Args:
+            image (numpy.ndarray): Original image.
+
+        Returns:
+            numpy.ndarray, image with perlin noise.
+        """
+        ori_dtype = image.dtype
+        _, chw, normalized, gray3dim, image = self._check(image)
+        w, h = image.shape[:2]
+        if self.auto_param:
+            self._set_auto_param(w, h)
+        if self.mode == 'circle':
+            mask = _circle_gradient_mask(image, self.color_start, self.color_end, self.scope, self.start_point)
+        else:
+            mask = _line_gradient_mask(image, self.start_point, self.color_start, self.color_end, mode=self.mode)
+
+        if self.pattern == 'light':
+            img_new = cv2.addWeighted(image, 1, mask, self.bright_rate, 0.0)
+        else:
+            img_new = cv2.addWeighted(image, self.bright_rate, mask, 1 - self.bright_rate, 0.0)
+        img_new = self._original_format(img_new, chw, normalized, gray3dim)
+        return img_new.astype(ori_dtype)

mindarmour/natural_robustness/image/natural_perturb.py

@@ -0,0 +1,159 @@
+# Copyright 2022 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Base class for image natural perturbation.
+"""
+import numpy as np
+
+from mindspore.dataset.vision.py_transforms_util import is_numpy, hwc_to_chw
+from mindarmour.utils.logger import LogUtil
+
+LOGGER = LogUtil.get_instance()
+TAG = 'Image Transformation'
+
+
+def _chw_to_hwc(img):
+    """
+    Transpose the input image; shape (C, H, W) to shape (H, W, C).
+
+    Args:
+        img (numpy.ndarray): Image to be converted.
+
+    Returns:
+        img (numpy.ndarray), Converted image.
+    """
+    if is_numpy(img):
+        return img.transpose(1, 2, 0).copy()
+    raise TypeError('img should be numpy.ndarray. Got {}'.format(type(img)))
+
+
+def _is_hwc(img):
+    """
+    Check if the input image is shape (H, W, C).
+
+    Args:
+        img (numpy.ndarray): Image to be checked.
+
+    Returns:
+        Bool, True if input is shape (H, W, C).
+    """
+    if is_numpy(img):
+        img_shape = np.shape(img)
+        if img_shape[2] == 3 and img_shape[1] > 3 and img_shape[0] > 3:
+            return True
+        return False
+    raise TypeError('img should be numpy.ndarray. Got {}'.format(type(img)))
+
+
+def _is_chw(img):
+    """
+    Check if the input image is shape (H, W, C).
+
+    Args:
+        img (numpy.ndarray): Image to be checked.
+
+    Returns:
+        Bool, True if input is shape (H, W, C).
+    """
+    if is_numpy(img):
+        img_shape = np.shape(img)
+        if img_shape[0] == 3 and img_shape[1] > 3 and img_shape[2] > 3:
+            return True
+        return False
+    raise TypeError('img should be numpy.ndarray. Got {}'.format(type(img)))
+
+
+def _is_rgb(img):
+    """
+    Check if the input image is RGB.
+
+    Args:
+        img (numpy.ndarray): Image to be checked.
+
+    Returns:
+        Bool, True if input is RGB.
+    """
+    if is_numpy(img):
+        img_shape = np.shape(img)
+        if len(np.shape(img)) == 3 and (img_shape[0] == 3 or img_shape[2] == 3):
+            return True
+        return False
+    raise TypeError('img should be numpy.ndarray. Got {}'.format(type(img)))
+
+
+def _is_normalized(img):
+    """
+    Check if the input image is normalized between 0 to 1.
+
+    Args:
+        img (numpy.ndarray): Image to be checked.
+
+    Returns:
+        Bool, True if input is normalized between 0 to 1.
+    """
+    if is_numpy(img):
+        minimal = np.min(img)
+        maximum = np.max(img)
+        if minimal >= 0 and maximum <= 1:
+            return True
+        return False
+    raise TypeError('img should be Numpy array. Got {}'.format(type(img)))
+
+
+class _NaturalPerturb:
+    """
+    The abstract base class for all image natural perturbation classes.
+    """
+
+    def __init__(self):
+        pass
+
+    def _check(self, image):
+        """ Check image format. If input image is RGB and its shape
+        is (C, H, W), it will be transposed to (H, W, C). If the value
+        of the image is not normalized , it will be rescaled between 0 to 255."""
+        rgb = _is_rgb(image)
+        chw = False
+        gray3dim = False
+        normalized = _is_normalized(image)
+        if rgb:
+            chw = _is_chw(image)
+            if chw:
+                image = _chw_to_hwc(image)
+            else:
+                image = image
+        else:
+            if len(np.shape(image)) == 3:
+                gray3dim = True
+                image = image[0]
+            else:
+                image = image
+        if normalized:
+            image = image * 255
+        return rgb, chw, normalized, gray3dim, np.uint8(image)
+
+    def _original_format(self, image, chw, normalized, gray3dim):
+        """ Return image with original format. """
+        if not is_numpy(image):
+            image = np.array(image)
+        if chw:
+            image = hwc_to_chw(image)
+        if normalized:
+            image = image / 255
+        if gray3dim:
+            image = np.expand_dims(image, 0)
+        return image
+
+    def __call__(self, image):
+        pass

mindarmour/natural_robustness/image/transformation.py

@@ -0,0 +1,365 @@
+# Copyright 2022 Huawei Technologies Co., Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Image transformation.
+"""
+import math
+import numpy as np
+import cv2
+
+from mindarmour.natural_robustness.image.natural_perturb import _NaturalPerturb
+from mindarmour.utils._check_param import check_param_multi_types, check_param_type, check_value_non_negative
+from mindarmour.utils.logger import LogUtil
+
+LOGGER = LogUtil.get_instance()
+TAG = 'Image Transformation'
+
+
+class Translate(_NaturalPerturb):
+    """
+    Translate an image.
+
+    Args:
+        x_bias (Union[int, float]): X-direction translation, x = x + x_bias*image_width. Suggested value range
+            in [-0.1, 0.1].
+        y_bias (Union[int, float]): Y-direction translation,  y = y + y_bias*image_length. Suggested value range
+            in [-0.1, 0.1].
+        auto_param (bool): Auto selected parameters. Selected parameters will preserve semantics of image.
+
+    Example:
+        >>> img = cv2.imread('1.png')
+        >>> img = np.array(img)
+        >>> x_bias = 0.1
+        >>> y_bias = 0.1
+        >>> trans = Translate(x_bias, y_bias)
+        >>> dst = trans(img)
+    """
+
+    def __init__(self, x_bias=0, y_bias=0, auto_param=False):
+        super(Translate, self).__init__()
+        self.x_bias = check_param_multi_types('x_bias', x_bias, [int, float])
+        self.y_bias = check_param_multi_types('y_bias', y_bias, [int, float])
+        if auto_param:
+            self.x_bias = np.random.uniform(-0.1, 0.1)
+            self.y_bias = np.random.uniform(-0.1, 0.1)
+
+    def __call__(self, image):
+        """
+        Transform the image.
+
+        Args:
+            image (numpy.ndarray): Original image to be transformed.
+
+        Returns:
+            numpy.ndarray, transformed image.
+        """
+        ori_dtype = image.dtype
+        _, chw, normalized, gray3dim, image = self._check(image)
+        h, w = image.shape[:2]
+        matrix = np.array([[1, 0, self.x_bias * w], [0, 1, self.y_bias * h]], dtype=np.float)
+        new_img = cv2.warpAffine(image, matrix, (w, h))
+        new_img = self._original_format(new_img, chw, normalized, gray3dim)
+        return new_img.astype(ori_dtype)
+
+
+class Scale(_NaturalPerturb):
+    """
+    Scale an image in the middle.
+
+    Args:
+        factor_x (Union[float, int]): Rescale in X-direction, x=factor_x*x. Suggested value range in [0.5, 1] and
+            abs(factor_y - factor_x) < 0.5.
+        factor_y (Union[float, int]): Rescale in Y-direction, y=factor_y*y. Suggested value range in [0.5, 1] and
+            abs(factor_y - factor_x) < 0.5.
+        auto_param (bool): Auto selected parameters. Selected parameters will preserve semantics of image.
+
+    Example:
+        >>> img = cv2.imread('1.png')
+        >>> img = np.array(img)
+        >>> factor_x = 0.7
+        >>> factor_y = 0.6
+        >>> trans = Scale(factor_x, factor_y)
+        >>> dst = trans(img)
+    """
+
+    def __init__(self, factor_x=1, factor_y=1, auto_param=False):
+        super(Scale, self).__init__()
+        self.factor_x = check_param_multi_types('factor_x', factor_x, [int, float])
+        self.factor_y = check_param_multi_types('factor_y', factor_y, [int, float])
+        auto_param = check_param_type('auto_param', auto_param, bool)
+        if auto_param:
+            self.factor_x = np.random.uniform(0.5, 1)
+            self.factor_y = np.random.uniform(0.5, 1)
+
+    def __call__(self, image):
+        """
+        Transform the image.
+
+        Args:
+            image (numpy.ndarray): Original image to be transformed.
+
+        Returns:
+            numpy.ndarray, transformed image.
+        """
+        ori_dtype = image.dtype
+        _, chw, normalized, gray3dim, image = self._check(image)
+        h, w = image.shape[:2]
+        matrix = np.array([[self.factor_x, 0, 0], [0, self.factor_y, 0]], dtype=np.float)
+        new_img = cv2.warpAffine(image, matrix, (w, h))
+        new_img = self._original_format(new_img, chw, normalized, gray3dim)
+        return new_img.astype(ori_dtype)
+
+
+class Shear(_NaturalPerturb):
+    """
+    Shear an image, for each pixel (x, y) in the sheared image, the new value is taken from a position
+    (x+factor_x*y, factor_y*x+y) in the origin image. Then the sheared image will be rescaled to fit original size.
+
+    Args:
+        factor (Union[float, int]): Shear rate in shear direction. Suggested value range in [0.05, 0.5].
+        direction (str): Direction of deformation. Optional value is 'vertical' or 'horizontal'.
+        auto_param (bool): Auto selected parameters. Selected parameters will preserve semantics of image.
+
+    Example:
+        >>> img = cv2.imread('1.png')
+        >>> img = np.array(img)
+        >>> factor = 0.2
+        >>> trans = Shear(factor, direction='horizontal')
+        >>> dst = trans(img)
+    """
+
+    def __init__(self, factor=0.2, direction='horizontal', auto_param=False):
+        super(Shear, self).__init__()
+        self.factor = check_param_multi_types('factor', factor, [int, float])
+        if direction not in ['horizontal', 'vertical']:
+            msg = "'direction must be in ['horizontal', 'vertical'], but got {}".format(direction)
+            raise ValueError(msg)
+        self.direction = direction
+        auto_param = check_param_type('auto_params', auto_param, bool)
+        if auto_param:
+            self.factor = np.random.uniform(0.05, 0.5)
+            self.direction = np.random.choice(['horizontal', 'vertical'])
+
+    def __call__(self, image):
+        """
+        Transform the image.
+
+        Args:
+            image (numpy.ndarray): Original image to be transformed.
+
+        Returns:
+            numpy.ndarray, transformed image.
+        """
+        ori_dtype = image.dtype
+        _, chw, normalized, gray3dim, image = self._check(image)
+        h, w = image.shape[:2]
+        if self.direction == 'horizontal':
+            matrix = np.array([[1, self.factor, 0], [0, 1, 0]], dtype=np.float)
+            nw = int(w + self.factor * h)
+            nh = h
+        else:
+            matrix = np.array([[1, 0, 0], [self.factor, 1, 0]], dtype=np.float)
+            nw = w
+            nh = int(h + self.factor * w)
+        new_img = cv2.warpAffine(image, matrix, (nw, nh))
+        new_img = cv2.resize(new_img, (w, h))
+        new_img = self._original_format(new_img, chw, normalized, gray3dim)
+        return new_img.astype(ori_dtype)
+
+
+class Rotate(_NaturalPerturb):
+    """
+    Rotate an image of counter clockwise around its center.
+
+    Args:
+        angle (Union[float, int]): Degrees of counter clockwise. Suggested value range in [-60, 60].
+        auto_param (bool): Auto selected parameters. Selected parameters will preserve semantics of image.
+
+    Example:
+        >>> img = cv2.imread('1.png')
+        >>> img = np.array(img)
+        >>> angle = 20
+        >>> trans = Rotate(angle)
+        >>> dst = trans(img)
+    """
+
+    def __init__(self, angle=20, auto_param=False):
+        super(Rotate, self).__init__()
+        self.angle = check_param_multi_types('angle', angle, [int, float])
+        auto_param = check_param_type('auto_param', auto_param, bool)
+        if auto_param:
+            self.angle = np.random.uniform(0, 360)
+
+    def __call__(self, image):
+        """
+        Transform the image.
+
+        Args:
+            image (numpy.ndarray): Original image to be transformed.
+
+        Returns:
+            numpy.ndarray, rotated image.
+        """
+        ori_dtype = image.dtype
+        _, chw, normalized, gray3dim, image = self._check(image)
+        h, w = image.shape[:2]
+        center = (w // 2, h // 2)
+        matrix = cv2.getRotationMatrix2D(center, -self.angle, 1.0)
+        cos = np.abs(matrix[0, 0])
+        sin = np.abs(matrix[0, 1])
+
+        # Calculate new edge after rotated
+        nw = int((h * sin) + (w * cos))
+        nh = int((h * cos) + (w * sin))
+        # Adjust move distance of rotate matrix.
+        matrix[0, 2] += (nw / 2) - center[0]
+        matrix[1, 2] += (nh / 2) - center[1]
+        rotate = cv2.warpAffine(image, matrix, (nw, nh))
+        rotate = cv2.resize(rotate, (w, h))
+        new_img = self._original_format(rotate, chw, normalized, gray3dim)
+        return new_img.astype(ori_dtype)
+
+
+class Perspective(_NaturalPerturb):
+    """
+    Perform perspective transformation on a given picture.
+
+    Args:
+        ori_pos (list): Four points in original image.
+        dst_pos (list): The point coordinates of the 4 points in ori_pos after perspective transformation.
+        auto_param (bool): Auto selected parameters. Selected parameters will preserve semantics of image.
+
+    Example:
+        >>> img = cv2.imread('1.png')
+        >>> img = np.array(img)
+        >>> ori_pos = [[0, 0], [0, 800], [800, 0], [800, 800]]
+        >>> dst_pos = [[50, 0], [0, 800], [780, 0], [800, 800]]
+        >>> trans = Perspective(ori_pos, dst_pos)
+        >>> dst = trans(img)
+    """
+
+    def __init__(self, ori_pos, dst_pos, auto_param=False):
+        super(Perspective, self).__init__()
+        ori_pos = check_param_type('ori_pos', ori_pos, list)
+        dst_pos = check_param_type('dst_pos', dst_pos, list)
+        self.ori_pos = np.float32(ori_pos)
+        self.dst_pos = np.float32(dst_pos)
+        self.auto_param = check_param_type('auto_param', auto_param, bool)
+
+    def _set_auto_param(self, w, h):
+        self.ori_pos = [[h * 0.25, w * 0.25], [h * 0.25, w * 0.75], [h * 0.75, w * 0.25], [h * 0.75, w * 0.75]]
+        self.dst_pos = [[np.random.uniform(0, h * 0.5), np.random.uniform(0, w * 0.5)],
+                        [np.random.uniform(0, h * 0.5), np.random.uniform(w * 0.5, w)],
+                        [np.random.uniform(h * 0.5, h), np.random.uniform(0, w * 0.5)],
+                        [np.random.uniform(h * 0.5, h), np.random.uniform(w * 0.5, w)]]
+        self.ori_pos = np.float32(self.ori_pos)
+        self.dst_pos = np.float32(self.dst_pos)
+
+    def __call__(self, image):
+        """
+        Transform the image.
+
+        Args:
+            image (numpy.ndarray): Original image to be transformed.
+
+        Returns:
+            numpy.ndarray, transformed image.
+        """
+        ori_dtype = image.dtype
+        _, chw, normalized, gray3dim, image = self._check(image)
+        h, w = image.shape[:2]
+        if self.auto_param:
+            self._set_auto_param(w, h)
+        matrix = cv2.getPerspectiveTransform(self.ori_pos, self.dst_pos)
+        new_img = cv2.warpPerspective(image, matrix, (w, h))
+        new_img = self._original_format(new_img, chw, normalized, gray3dim)
+        return new_img.astype(ori_dtype)
+
+
+class Curve(_NaturalPerturb):
+    """
+    Curve picture using sin method.
+
+    Args:
+        curves (union[float, int]): Divide width to curves of `2*math.pi`, which means how many curve cycles. Suggested
+            value range in [0.1. 5].
+        depth (union[float, int]): Amplitude of sin method. Suggested value not exceed 1/10 of the length of the
+            picture.
+        mode (str): Direction of deformation. Optional value is 'vertical' or 'horizontal'.
+        auto_param (bool): Auto selected parameters. Selected parameters will preserve semantics of image.
+
+    Examples:
+        >>> img = cv2.imread('x.png')
+        >>> curves =1
+        >>> depth = 10
+        >>> trans = Curve(curves, depth, mode='vertical')
+        >>> img_new = trans(img)
+    """
+
+    def __init__(self, curves=3, depth=10, mode='vertical', auto_param=False):
+        super(Curve).__init__()
+        self.curves = check_value_non_negative('curves', curves)
+        self.depth = check_value_non_negative('depth', depth)
+        if mode in ['vertical', 'horizontal']:
+            self.mode = mode
+        else:
+            msg = "Value of param mode must be in ['vertical', 'horizontal']"
+            LOGGER.error(TAG, msg)
+            raise ValueError(msg)
+        self.auto_param = check_param_type('auto_param', auto_param, bool)
+
+    def _set_auto_params(self, height, width):
+        if self.auto_param:
+            self.curves = np.random.uniform(1, 5)
+            self.mode = np.random.choice(['vertical', 'horizontal'])
+            if self.mode == 'vertical':
+                self.depth = np.random.uniform(1, 0.1 * width)
+            else:
+                self.depth = np.random.uniform(1, 0.1 * height)
+
+    def __call__(self, image):
+        """
+        Curve picture using sin method.
+
+        Args:
+            image (numpy.ndarray): Original image.
+
+        Returns:
+            numpy.ndarray, curved image.
+        """
+        ori_dtype = image.dtype
+        _, chw, normalized, gray3dim, image = self._check(image)
+        shape = image.shape
+        height, width = shape[:2]
+        if self.mode == 'vertical':
+            if len(shape) == 3:
+                image = np.transpose(image, [1, 0, 2])
+            else:
+                image = np.transpose(image, [1, 0])
+        src_x = np.zeros((height, width), np.float32)
+        src_y = np.zeros((height, width), np.float32)
+
+        for y in range(height):
+            for x in range(width):
+                src_x[y, x] = x
+                src_y[y, x] = y + self.depth * math.sin(x / (width / self.curves / 2 / math.pi))
+        img_new = cv2.remap(image, src_x, src_y, cv2.INTER_LINEAR)
+
+        if self.mode == 'vertical':
+            if len(shape) == 3:
+                img_new = np.transpose(img_new, [1, 0, 2])
+            else:
+                img_new = np.transpose(image, [1, 0])
+        new_img = self._original_format(img_new, chw, normalized, gray3dim)
+        return new_img.astype(ori_dtype)

setup.py

@@ -132,6 +132,7 @@ setup(
         'pillow >= 6.2.0',
         'scikit-learn >= 0.23.1',
         'easydict >= 1.9',
+        'opencv-python >= 4.1.2.30',
     ],
     classifiers=[
         'License :: OSI Approved :: Apache Software License'