corruptions.py

# -*- coding: utf-8 -*-

import numpy as np
from PIL import Image

# /////////////// Corruption Helpers ///////////////

import skimage as sk
from skimage.filters import gaussian
from io import BytesIO
from wand.image import Image as WandImage
from wand.api import library as wandlibrary
import wand.color as WandColor
import ctypes
from PIL import Image as PILImage
import cv2
from scipy.ndimage import zoom as scizoom
from scipy.ndimage.interpolation import map_coordinates
import warnings
import os
from pkg_resources import resource_filename

warnings.simplefilter("ignore", UserWarning)


def disk(radius, alias_blur=0.1, dtype=np.float32):
    if radius <= 8:
        L = np.arange(-8, 8 + 1)
        ksize = (3, 3)
    else:
        L = np.arange(-radius, radius + 1)
        ksize = (5, 5)
    X, Y = np.meshgrid(L, L)
    aliased_disk = np.array((X ** 2 + Y ** 2) <= radius ** 2, dtype=dtype)
    aliased_disk /= np.sum(aliased_disk)

    # supersample disk to antialias
    return cv2.GaussianBlur(aliased_disk, ksize=ksize, sigmaX=alias_blur)


# Tell Python about the C method
wandlibrary.MagickMotionBlurImage.argtypes = (ctypes.c_void_p,  # wand
                                              ctypes.c_double,  # radius
                                              ctypes.c_double,  # sigma
                                              ctypes.c_double)  # angle


# Extend wand.image.Image class to include method signature
class MotionImage(WandImage):
    def motion_blur(self, radius=0.0, sigma=0.0, angle=0.0):
        wandlibrary.MagickMotionBlurImage(self.wand, radius, sigma, angle)


# modification of https://github.com/FLHerne/mapgen/blob/master/diamondsquare.py
def plasma_fractal(mapsize=256, wibbledecay=3):
    """
    Generate a heightmap using diamond-square algorithm.
    Return square 2d array, side length 'mapsize', of floats in range 0-255.
    'mapsize' must be a power of two.
    """
    assert (mapsize & (mapsize - 1) == 0)
    maparray = np.empty((mapsize, mapsize), dtype=np.float_)
    maparray[0, 0] = 0
    stepsize = mapsize
    wibble = 100

    def wibbledmean(array):
        return array / 4 + wibble * np.random.uniform(-wibble, wibble, array.shape)

    def fillsquares():
        """For each square of points stepsize apart,
           calculate middle value as mean of points + wibble"""
        cornerref = maparray[0:mapsize:stepsize, 0:mapsize:stepsize]
        squareaccum = cornerref + np.roll(cornerref, shift=-1, axis=0)
        squareaccum += np.roll(squareaccum, shift=-1, axis=1)
        maparray[stepsize // 2:mapsize:stepsize,
        stepsize // 2:mapsize:stepsize] = wibbledmean(squareaccum)

    def filldiamonds():
        """For each diamond of points stepsize apart,
           calculate middle value as mean of points + wibble"""
        mapsize = maparray.shape[0]
        drgrid = maparray[stepsize // 2:mapsize:stepsize, stepsize // 2:mapsize:stepsize]
        ulgrid = maparray[0:mapsize:stepsize, 0:mapsize:stepsize]
        ldrsum = drgrid + np.roll(drgrid, 1, axis=0)
        lulsum = ulgrid + np.roll(ulgrid, -1, axis=1)
        ltsum = ldrsum + lulsum
        maparray[0:mapsize:stepsize, stepsize // 2:mapsize:stepsize] = wibbledmean(ltsum)
        tdrsum = drgrid + np.roll(drgrid, 1, axis=1)
        tulsum = ulgrid + np.roll(ulgrid, -1, axis=0)
        ttsum = tdrsum + tulsum
        maparray[stepsize // 2:mapsize:stepsize, 0:mapsize:stepsize] = wibbledmean(ttsum)

    while stepsize >= 2:
        fillsquares()
        filldiamonds()
        stepsize //= 2
        wibble /= wibbledecay

    maparray -= maparray.min()
    return maparray / maparray.max()


def clipped_zoom(img, zoom_factor):
    h = img.shape[0]
    # ceil crop height(= crop width)
    ch = int(np.ceil(h / float(zoom_factor)))

    top = (h - ch) // 2
    img = scizoom(img[top:top + ch, top:top + ch], (zoom_factor, zoom_factor, 1), order=1)
    # trim off any extra pixels
    trim_top = (img.shape[0] - h) // 2

    return img[trim_top:trim_top + h, trim_top:trim_top + h]


# /////////////// End Corruption Helpers ///////////////


# /////////////// Corruptions ///////////////

def gaussian_noise(x, severity=1):
    c = [.08, .12, 0.18, 0.24, 0.30][severity - 1]

    x = np.array(x) / 255.
    return np.clip(x + np.random.normal(size=x.shape, scale=c), 0, 1) * 255


def shot_noise(x, severity=1):
    c = [60, 29, 15, 8, 5][severity - 1]

    x = np.array(x) / 255.
    return np.clip(np.random.poisson(x * c) / float(c), 0, 1) * 255


def impulse_noise(x, severity=1):
    c = [.03, .06, .09, 0.17, 0.27][severity - 1]

    x = sk.util.random_noise(np.array(x) / 255., mode='s&p', amount=c)
    return np.clip(x, 0, 1) * 255


def speckle_noise(x, severity=1):
    c = [.15, .2, 0.35, 0.45, 0.6][severity - 1]

    x = np.array(x) / 255.
    return np.clip(x + x * np.random.normal(size=x.shape, scale=c), 0, 1) * 255


def fgsm(x, source_net, severity=1):
    c = [8, 16, 32, 64, 128][severity - 1]

    x = V(x, requires_grad=True)
    logits = source_net(x)
    source_net.zero_grad()
    loss = F.cross_entropy(logits, V(logits.data.max(1)[1].squeeze_()), size_average=False)
    loss.backward()

    return standardize(torch.clamp(unstandardize(x.data) + c / 255. * unstandardize(torch.sign(x.grad.data)), 0, 1))


def gaussian_blur(x, severity=1):
    c = [1, 1.8, 2.6, 3.4, 4.0][severity - 1]

    x = gaussian(np.array(x) / 255., sigma=c, multichannel=True)
    return np.clip(x, 0, 1) * 255


def glass_blur(x, severity=1):
    # sigma, max_delta, iterations
    c = [(0.7, 1, 2), (0.9, 2, 1), (1, 2, 3), (1.1, 3, 2), (1.5, 4, 2)][severity - 1]

    x = np.uint8(gaussian(np.array(x) / 255., sigma=c[0], multichannel=True) * 255)

    # locally shuffle pixels
    for i in range(c[2]):
        for h in range(224 - c[1], c[1], -1):
            for w in range(224 - c[1], c[1], -1):
                dx, dy = np.random.randint(-c[1], c[1], size=(2,))
                h_prime, w_prime = h + dy, w + dx
                # swap
                x[h, w], x[h_prime, w_prime] = x[h_prime, w_prime], x[h, w]

    return np.clip(gaussian(x / 255., sigma=c[0], multichannel=True), 0, 1) * 255


def defocus_blur(x, severity=1):
    c = [(1.5, 0.1), (2, 0.2), (2, 0.3), (2.5, 0.4), (3, 0.4)][severity - 1]

    c = tuple([item/48*x.shape[0] for item in c])
    
    x = np.array(x) / 255.
    kernel = disk(radius=c[0], alias_blur=c[1])

    channels = []
    for d in range(x.shape[2]):
        channels.append(cv2.filter2D(x[:, :, d], -1, kernel))
    channels = np.array(channels).transpose((1, 2, 0))  # 3x224x224 -> 224x224x3
    
    return (np.clip(channels, 0, 1) * 255).astype(np.uint8)


def motion_blur(x, severity=1):
    c = [(10, 3), (15, 5), (15, 8), (15, 12), (20, 15)][severity - 1]
    c = tuple([item/48*x.shape[0] for item in c])
    if len(x.shape)==3 and x.shape[2]==1:
        x = np.squeeze(x,2)

    x = Image.fromarray(x[...,[2,1,0]])

    output = BytesIO()
    x.save(output, format='PNG')
    x = MotionImage(blob=output.getvalue())

    x.motion_blur(radius=c[0]//3, sigma=c[1]//3,
        angle=np.random.uniform(-45, 45))

    x = cv2.imdecode(np.fromstring(x.make_blob(), np.uint8),
                     cv2.IMREAD_UNCHANGED)

    if len(x.shape)==2:
        x = np.expand_dims(x,2)
    return np.clip(x, 0, 255)

def custom_motion_blur(x, radius, sigma, angle):
    if len(x.shape)==3 and x.shape[2]==1:
        x = np.squeeze(x,2)

    x = Image.fromarray(x[...,[2,1,0]])

    output = BytesIO()
    x.save(output, format='PNG')
    x = MotionImage(blob=output.getvalue())

    x.motion_blur(radius, sigma, angle)

    x = cv2.imdecode(np.fromstring(x.make_blob(), np.uint8),
                     cv2.IMREAD_UNCHANGED)

    if len(x.shape)==2:
        x = np.expand_dims(x,2)
    return np.clip(x, 0, 255)


def zoom_blur(x, severity=1):
    c = [np.arange(1, 1.11, 0.01),
         np.arange(1, 1.18, 0.01),
         np.arange(1, 1.26, 0.02),
         np.arange(1, 1.32, 0.02),
         np.arange(1, 1.40, 0.03)][severity - 1]

    x = (np.array(x) / 255.).astype(np.float32)
    out = np.zeros_like(x)
    for zoom_factor in c:
        out += clipped_zoom(x, zoom_factor)

    x = (x + out) / (len(c) + 1)
    return np.clip(x, 0, 1) * 255


def fog(x, severity=1):
    c = [(1.5, 2), (2., 2), (2.5, 1.7), (2.5, 1.5), (3., 1.4)][severity - 1]

    x = np.array(x) / 255.
    max_val = x.max()
    x += c[0] * plasma_fractal(wibbledecay=c[1])[:224, :224][..., np.newaxis]
    return np.clip(x * max_val / (max_val + c[0]), 0, 1) * 255


def frost(x, severity=1):
    c = [(1, 0.4),
         (0.8, 0.6),
         (0.7, 0.7),
         (0.65, 0.7),
         (0.6, 0.75)][severity - 1]
    idx = np.random.randint(5)
    filename = [resource_filename(__name__, 'frost/frost1.png'),
                resource_filename(__name__, 'frost/frost2.png'),
                resource_filename(__name__, 'frost/frost3.png'),
                resource_filename(__name__, 'frost/frost4.jpg'),
                resource_filename(__name__, 'frost/frost5.jpg'),
                resource_filename(__name__, 'frost/frost6.jpg')][idx]
    frost = cv2.imread(filename)
    # randomly crop and convert to rgb
    x_start, y_start = np.random.randint(0, frost.shape[0] - 224), np.random.randint(0, frost.shape[1] - 224)
    frost = frost[x_start:x_start + 224, y_start:y_start + 224][..., [2, 1, 0]]

    return np.clip(c[0] * np.array(x) + c[1] * frost, 0, 255)


def snow(x, severity=1):
    c = [(0.1, 0.3, 3, 0.5, 10, 4, 0.8),
         (0.2, 0.3, 2, 0.5, 12, 4, 0.7),
         (0.55, 0.3, 4, 0.9, 12, 8, 0.7),
         (0.55, 0.3, 4.5, 0.85, 12, 8, 0.65),
         (0.55, 0.3, 2.5, 0.85, 12, 12, 0.55)][severity - 1]

    x = np.array(x, dtype=np.float32) / 255.
    snow_layer = np.random.normal(size=x.shape[:2], loc=c[0], scale=c[1])  # [:2] for monochrome

    snow_layer = clipped_zoom(snow_layer[..., np.newaxis], c[2])
    snow_layer[snow_layer < c[3]] = 0

    snow_layer = PILImage.fromarray((np.clip(snow_layer.squeeze(), 0, 1) * 255).astype(np.uint8), mode='L')
    output = BytesIO()
    snow_layer.save(output, format='PNG')
    snow_layer = MotionImage(blob=output.getvalue())

    snow_layer.motion_blur(radius=c[4], sigma=c[5], angle=np.random.uniform(-135, -45))

    snow_layer = cv2.imdecode(np.fromstring(snow_layer.make_blob(), np.uint8),
                              cv2.IMREAD_UNCHANGED) / 255.
    snow_layer = snow_layer[..., np.newaxis]

    x = c[6] * x + (1 - c[6]) * np.maximum(x, cv2.cvtColor(x, cv2.COLOR_RGB2GRAY).reshape(224, 224, 1) * 1.5 + 0.5)
    return np.clip(x + snow_layer + np.rot90(snow_layer, k=2), 0, 1) * 255


def spatter(x, severity=1):
    iscolor = len(x.shape)>2 and x.shape[2] > 1
    c = [(0.65, 0.3, 4, 0.69, 0.6, 0),
         (0.65, 0.3, 3, 0.68, 0.6, 0),
         (0.65, 0.3, 2, 0.68, 0.5, 0),
         (0.65, 0.3, 1, 0.65, 1.5, 1),
         (0.67, 0.4, 1, 0.65, 1.5, 1)][severity - 1]
    x = np.array(x, dtype=np.float32) / 255.

    liquid_layer = np.random.normal(size=x.shape[:2], loc=c[0], scale=c[1])

    liquid_layer = gaussian(liquid_layer, sigma=c[2])
    liquid_layer[liquid_layer < c[3]] = 0
    if c[5] == 0:
        liquid_layer = (liquid_layer * 255).astype(np.uint8)
        dist = 255 - cv2.Canny(liquid_layer, 50, 150)
        dist = cv2.distanceTransform(dist, cv2.DIST_L2, 5)
        _, dist = cv2.threshold(dist, 20, 20, cv2.THRESH_TRUNC)
        dist = cv2.blur(dist, (3, 3)).astype(np.uint8)
        dist = cv2.equalizeHist(dist)
        ker = np.array([[-2, -1, 0], [-1, 1, 1], [0, 1, 2]])
        dist = cv2.filter2D(dist, cv2.CV_8U, ker)
        dist = cv2.blur(dist, (3, 3)).astype(np.float32)

        m = cv2.cvtColor(liquid_layer * dist, cv2.COLOR_GRAY2BGRA)
        m /= np.max(m, axis=(0, 1))
        m *= c[4]

        # water is pale turqouise
        color = np.concatenate((175 / 255. * np.ones_like(m[..., :1]),
                                238 / 255. * np.ones_like(m[..., :1]),
                                238 / 255. * np.ones_like(m[..., :1])), axis=2)
        
        color = cv2.cvtColor(color, cv2.COLOR_BGR2BGRA)
        if len(x.shape)>2 and x.shape[2] > 1:
            x = cv2.cvtColor(x, cv2.COLOR_BGR2BGRA)

        x = np.clip(x + m * color, 0, 1) * 255
        if iscolor: 
            return cv2.cvtColor(x, cv2.COLOR_BGRA2BGR)
        else:
            return cv2.cvtColor(x, cv2.COLOR_BGRA2GRAY)
    else:
        m = np.where(liquid_layer > c[3], 1, 0)
        m = gaussian(m.astype(np.float32), sigma=c[4])
        m[m < 0.8] = 0

        # mud brown
        color = np.concatenate((63 / 255. * np.ones_like(x[..., :1]),
                                42 / 255. * np.ones_like(x[..., :1]),
                                20 / 255. * np.ones_like(x[..., :1])), axis=2)

        color *= m[..., np.newaxis]
        x *= (1 - m[..., np.newaxis])

        x = np.clip(x + color, 0, 1) * 255
        
        if iscolor: 
            return x
        else:
            return cv2.cvtColor(x, cv2.COLOR_BGR2GRAY)

def contrast_plus(x, severity=1):
    c = [1.5, 1.9, 2.6, 3.3, 5.0][severity - 1]

    x = np.array(x) / 255.
    means = np.mean(x, axis=(0, 1), keepdims=True)
    return np.clip((x - means) * c + means, 0, 1) * 255

def contrast(x, severity=1):
    c = [.4, .33, .24, .16, .1][severity - 1]

    x = np.array(x) / 255.
    means = np.mean(x, axis=(0, 1), keepdims=True)
    return np.clip((x - means) * c + means, 0, 1) * 255


def brightness_plus(x, severity=1):
    c = [.1, .2, .3, .4, .5][severity - 1]
    return brightness(x, c)
def brightness_minus(x, severity=1):
    c = [.1, .2, .3, .4, .5][severity - 1]
    return brightness(x, -c)

def brightness(x, c):
    x = np.array(x) / 255.
    if len(x.shape)>2 and x.shape[2]>1:
        x = sk.color.rgb2hsv(x)
        x[:, :, 2] = np.clip(x[:, :, 2] + c, 0, 1)
        x = sk.color.hsv2rgb(x)
    else:
        x = np.clip(x + c, 0, 1)
    
    return np.clip(x, 0, 1) * 255


def saturate(x, severity=1):
    c = [(0.3, 0), (0.1, 0), (2, 0), (5, 0.1), (20, 0.2)][severity - 1]

    x = np.array(x) / 255.
    x = sk.color.rgb2hsv(x)
    x[:, :, 1] = np.clip(x[:, :, 1] * c[0] + c[1], 0, 1)
    x = sk.color.hsv2rgb(x)

    return np.clip(x, 0, 1) * 255


def jpeg_compression(x, severity=1):
    c = [25, 18, 15, 10, 7][severity - 1]

    if len(x.shape)==3 and x.shape[2]==1:
        x = np.squeeze(x,2)
    x = Image.fromarray(x)

    output = BytesIO()
    x.save(output, 'JPEG', quality=c)
    x = PILImage.open(output)
    x = np.array(x)
    if len(x.shape)==2:
        x = np.expand_dims(x,2)
    return x


def pixelate(x, severity=1):
    c = [0.6, 0.5, 0.41, 0.3, 0.25][severity - 1]
    original_shape = x.shape

    if len(x.shape)==3 and x.shape[2]==1:
        x = np.squeeze(x,2)
    x = Image.fromarray(x)
    x = x.resize((int(original_shape[0] * c), int(original_shape[1] * c)), PILImage.BOX)
    x = x.resize((original_shape[0], original_shape[1]), PILImage.BOX)
    x = np.array(x)
    if len(x.shape)==2:
        x = np.expand_dims(x,2)

    return x


# mod of https://gist.github.com/erniejunior/601cdf56d2b424757de5
def elastic_transform(image, severity=1):
    c = [(image.shape[0] * 2, image.shape[0] * 0.7, image.shape[0] * 0.1),
         (image.shape[0] * 2, image.shape[0] * 0.08, image.shape[0] * 0.2),
         (image.shape[0] * 0.05, image.shape[0] * 0.01, image.shape[0] * 0.02),
         (image.shape[0] * 0.07, image.shape[0] * 0.01, image.shape[0] * 0.02),
         (image.shape[0] * 0.12, image.shape[0] * 0.01, image.shape[0] * 0.02)][severity - 1]

    image = np.array(image, dtype=np.float32) / 255.
    shape = image.shape

    shape_size = shape[:2]

    # random affine
    '''
    center_square = np.float32(shape_size) // 2
    square_size = min(shape_size) // 3
    pts1 = np.float32([center_square + square_size,
                       [center_square[0] + square_size, center_square[1] - square_size],
                       center_square - square_size])
    pts2 = pts1 + np.random.uniform(-c[2], c[2], size=pts1.shape).astype(np.float32)
    M = cv2.getAffineTransform(pts1, pts2)
    image = cv2.warpAffine(image, M, shape_size[::-1], borderMode=cv2.BORDER_REFLECT_101)
    if len(image.shape)<3:
        image = np.expand_dims(image,2)
    '''
    dx = (gaussian(np.random.uniform(-1, 1, size=shape[:2]),
                   c[1], mode='reflect', truncate=3) * c[0]).astype(np.float32)
    dy = (gaussian(np.random.uniform(-1, 1, size=shape[:2]),
                   c[1], mode='reflect', truncate=3) * c[0]).astype(np.float32)
    dx, dy = dx[..., np.newaxis], dy[..., np.newaxis]

    x, y, z = np.meshgrid(np.arange(shape[1]), np.arange(shape[0]), np.arange(shape[2]))
    indices = np.reshape(y + dy, (-1, 1)), np.reshape(x + dx, (-1, 1)), np.reshape(z, (-1, 1))

    return np.clip(map_coordinates(image, indices, order=1, mode='reflect').reshape(shape), 0, 1) * 255


# /////////////// End Corruptions ///////////////