captchaboy/dfff4444
0
1import math2import numbers3import random4 5import cv26import numpy as np7from PIL import Image8from torchvision import transforms9from torchvision.transforms import Compose10 11 12def sample_asym(magnitude, size=None):13 return np.random.beta(1, 4, size) * magnitude14 15def sample_sym(magnitude, size=None):16 return (np.random.beta(4, 4, size=size) - 0.5) * 2 * magnitude17 18def sample_uniform(low, high, size=None):19 return np.random.uniform(low, high, size=size)20 21def get_interpolation(type='random'):22 if type == 'random':23 choice = [cv2.INTER_NEAREST, cv2.INTER_LINEAR, cv2.INTER_CUBIC, cv2.INTER_AREA]24 interpolation = choice[random.randint(0, len(choice)-1)]25 elif type == 'nearest': interpolation = cv2.INTER_NEAREST26 elif type == 'linear': interpolation = cv2.INTER_LINEAR27 elif type == 'cubic': interpolation = cv2.INTER_CUBIC28 elif type == 'area': interpolation = cv2.INTER_AREA29 else: raise TypeError('Interpolation types only nearest, linear, cubic, area are supported!')30 return interpolation31 32class CVRandomRotation(object):33 def __init__(self, degrees=15):34 assert isinstance(degrees, numbers.Number), "degree should be a single number."35 assert degrees >= 0, "degree must be positive."36 self.degrees = degrees37 38 @staticmethod39 def get_params(degrees):40 return sample_sym(degrees)41 42 def __call__(self, img):43 angle = self.get_params(self.degrees)44 src_h, src_w = img.shape[:2]45 M = cv2.getRotationMatrix2D(center=(src_w/2, src_h/2), angle=angle, scale=1.0)46 abs_cos, abs_sin = abs(M[0,0]), abs(M[0,1])47 dst_w = int(src_h * abs_sin + src_w * abs_cos)48 dst_h = int(src_h * abs_cos + src_w * abs_sin)49 M[0, 2] += (dst_w - src_w)/250 M[1, 2] += (dst_h - src_h)/251 52 flags = get_interpolation()53 return cv2.warpAffine(img, M, (dst_w, dst_h), flags=flags, borderMode=cv2.BORDER_REPLICATE)54 55class CVRandomAffine(object):56 def __init__(self, degrees, translate=None, scale=None, shear=None):57 assert isinstance(degrees, numbers.Number), "degree should be a single number."58 assert degrees >= 0, "degree must be positive."59 self.degrees = degrees60 61 if translate is not None:62 assert isinstance(translate, (tuple, list)) and len(translate) == 2, \63 "translate should be a list or tuple and it must be of length 2."64 for t in translate:65 if not (0.0 <= t <= 1.0):66 raise ValueError("translation values should be between 0 and 1")67 self.translate = translate68 69 if scale is not None:70 assert isinstance(scale, (tuple, list)) and len(scale) == 2, \71 "scale should be a list or tuple and it must be of length 2."72 for s in scale:73 if s <= 0:74 raise ValueError("scale values should be positive")75 self.scale = scale76 77 if shear is not None:78 if isinstance(shear, numbers.Number):79 if shear < 0:80 raise ValueError("If shear is a single number, it must be positive.")81 self.shear = [shear]82 else:83 assert isinstance(shear, (tuple, list)) and (len(shear) == 2), \84 "shear should be a list or tuple and it must be of length 2."85 self.shear = shear86 else:87 self.shear = shear88 89 def _get_inverse_affine_matrix(self, center, angle, translate, scale, shear):90 # https://github.com/pytorch/vision/blob/v0.4.0/torchvision/transforms/functional.py#L71791 from numpy import sin, cos, tan92 93 if isinstance(shear, numbers.Number):94 shear = [shear, 0]95 96 if not isinstance(shear, (tuple, list)) and len(shear) == 2:97 raise ValueError(98 "Shear should be a single value or a tuple/list containing " +99 "two values. Got {}".format(shear))100 101 rot = math.radians(angle)102 sx, sy = [math.radians(s) for s in shear]103 104 cx, cy = center105 tx, ty = translate106 107 # RSS without scaling108 a = cos(rot - sy) / cos(sy)109 b = -cos(rot - sy) * tan(sx) / cos(sy) - sin(rot)110 c = sin(rot - sy) / cos(sy)111 d = -sin(rot - sy) * tan(sx) / cos(sy) + cos(rot)112 113 # Inverted rotation matrix with scale and shear114 # det([[a, b], [c, d]]) == 1, since det(rotation) = 1 and det(shear) = 1115 M = [d, -b, 0,116 -c, a, 0]117 M = [x / scale for x in M]118 119 # Apply inverse of translation and of center translation: RSS^-1 * C^-1 * T^-1120 M[2] += M[0] * (-cx - tx) + M[1] * (-cy - ty)121 M[5] += M[3] * (-cx - tx) + M[4] * (-cy - ty)122 123 # Apply center translation: C * RSS^-1 * C^-1 * T^-1124 M[2] += cx125 M[5] += cy126 return M127 128 @staticmethod129 def get_params(degrees, translate, scale_ranges, shears, height): 130 angle = sample_sym(degrees)131 if translate is not None:132 max_dx = translate[0] * height133 max_dy = translate[1] * height134 translations = (np.round(sample_sym(max_dx)), np.round(sample_sym(max_dy)))135 else:136 translations = (0, 0)137 138 if scale_ranges is not None:139 scale = sample_uniform(scale_ranges[0], scale_ranges[1])140 else:141 scale = 1.0142 143 if shears is not None:144 if len(shears) == 1:145 shear = [sample_sym(shears[0]), 0.]146 elif len(shears) == 2:147 shear = [sample_sym(shears[0]), sample_sym(shears[1])]148 else:149 shear = 0.0150 151 return angle, translations, scale, shear152 153 154 def __call__(self, img):155 src_h, src_w = img.shape[:2]156 angle, translate, scale, shear = self.get_params(157 self.degrees, self.translate, self.scale, self.shear, src_h)158 159 M = self._get_inverse_affine_matrix((src_w/2, src_h/2), angle, (0, 0), scale, shear)160 M = np.array(M).reshape(2,3)161 162 startpoints = [(0, 0), (src_w - 1, 0), (src_w - 1, src_h - 1), (0, src_h - 1)]163 project = lambda x, y, a, b, c: int(a*x + b*y + c)164 endpoints = [(project(x, y, *M[0]), project(x, y, *M[1])) for x, y in startpoints]165 166 rect = cv2.minAreaRect(np.array(endpoints))167 bbox = cv2.boxPoints(rect).astype(dtype=np.int)168 max_x, max_y = bbox[:, 0].max(), bbox[:, 1].max()169 min_x, min_y = bbox[:, 0].min(), bbox[:, 1].min()170 171 dst_w = int(max_x - min_x)172 dst_h = int(max_y - min_y)173 M[0, 2] += (dst_w - src_w) / 2174 M[1, 2] += (dst_h - src_h) / 2175 176 # add translate177 dst_w += int(abs(translate[0]))178 dst_h += int(abs(translate[1]))179 if translate[0] < 0: M[0, 2] += abs(translate[0])180 if translate[1] < 0: M[1, 2] += abs(translate[1])181 182 flags = get_interpolation()183 return cv2.warpAffine(img, M, (dst_w , dst_h), flags=flags, borderMode=cv2.BORDER_REPLICATE)184 185class CVRandomPerspective(object):186 def __init__(self, distortion=0.5):187 self.distortion = distortion188 189 def get_params(self, width, height, distortion):190 offset_h = sample_asym(distortion * height / 2, size=4).astype(dtype=np.int)191 offset_w = sample_asym(distortion * width / 2, size=4).astype(dtype=np.int)192 topleft = ( offset_w[0], offset_h[0])193 topright = (width - 1 - offset_w[1], offset_h[1])194 botright = (width - 1 - offset_w[2], height - 1 - offset_h[2])195 botleft = ( offset_w[3], height - 1 - offset_h[3])196 197 startpoints = [(0, 0), (width - 1, 0), (width - 1, height - 1), (0, height - 1)]198 endpoints = [topleft, topright, botright, botleft]199 return np.array(startpoints, dtype=np.float32), np.array(endpoints, dtype=np.float32)200 201 def __call__(self, img):202 height, width = img.shape[:2]203 startpoints, endpoints = self.get_params(width, height, self.distortion)204 M = cv2.getPerspectiveTransform(startpoints, endpoints)205 206 # TODO: more robust way to crop image207 rect = cv2.minAreaRect(endpoints)208 bbox = cv2.boxPoints(rect).astype(dtype=np.int)209 max_x, max_y = bbox[:, 0].max(), bbox[:, 1].max()210 min_x, min_y = bbox[:, 0].min(), bbox[:, 1].min()211 min_x, min_y = max(min_x, 0), max(min_y, 0)212 213 flags = get_interpolation() 214 img = cv2.warpPerspective(img, M, (max_x, max_y), flags=flags, borderMode=cv2.BORDER_REPLICATE)215 img = img[min_y:, min_x:]216 return img217 218class CVRescale(object):219 220 def __init__(self, factor=4, base_size=(128, 512)):221 """ Define image scales using gaussian pyramid and rescale image to target scale.222 223 Args:224 factor: the decayed factor from base size, factor=4 keeps target scale by default.225 base_size: base size the build the bottom layer of pyramid226 """227 if isinstance(factor, numbers.Number):228 self.factor = round(sample_uniform(0, factor))229 elif isinstance(factor, (tuple, list)) and len(factor) == 2:230 self.factor = round(sample_uniform(factor[0], factor[1]))231 else:232 raise Exception('factor must be number or list with length 2')233 # assert factor is valid234 self.base_h, self.base_w = base_size[:2]235 236 def __call__(self, img):237 if self.factor == 0: return img 238 src_h, src_w = img.shape[:2]239 cur_w, cur_h = self.base_w, self.base_h 240 scale_img = cv2.resize(img, (cur_w, cur_h), interpolation=get_interpolation())241 for _ in range(self.factor): 242 scale_img = cv2.pyrDown(scale_img)243 scale_img = cv2.resize(scale_img, (src_w, src_h), interpolation=get_interpolation())244 return scale_img245 246class CVGaussianNoise(object):247 def __init__(self, mean=0, var=20):248 self.mean = mean249 if isinstance(var, numbers.Number):250 self.var = max(int(sample_asym(var)), 1)251 elif isinstance(var, (tuple, list)) and len(var) == 2:252 self.var = int(sample_uniform(var[0], var[1]))253 else:254 raise Exception('degree must be number or list with length 2')255 256 def __call__(self, img):257 noise = np.random.normal(self.mean, self.var**0.5, img.shape)258 img = np.clip(img + noise, 0, 255).astype(np.uint8)259 return img260 261class CVMotionBlur(object):262 def __init__(self, degrees=12, angle=90):263 if isinstance(degrees, numbers.Number):264 self.degree = max(int(sample_asym(degrees)), 1)265 elif isinstance(degrees, (tuple, list)) and len(degrees) == 2:266 self.degree = int(sample_uniform(degrees[0], degrees[1]))267 else:268 raise Exception('degree must be number or list with length 2')269 self.angle = sample_uniform(-angle, angle)270 271 def __call__(self, img):272 M = cv2.getRotationMatrix2D((self.degree // 2, self.degree // 2), self.angle, 1)273 motion_blur_kernel = np.zeros((self.degree, self.degree))274 motion_blur_kernel[self.degree // 2, :] = 1275 motion_blur_kernel = cv2.warpAffine(motion_blur_kernel, M, (self.degree, self.degree))276 motion_blur_kernel = motion_blur_kernel / self.degree277 img = cv2.filter2D(img, -1, motion_blur_kernel)278 img = np.clip(img, 0, 255).astype(np.uint8)279 return img280 281class CVGeometry(object):282 def __init__(self, degrees=15, translate=(0.3, 0.3), scale=(0.5, 2.), 283 shear=(45, 15), distortion=0.5, p=0.5):284 self.p = p285 type_p = random.random()286 if type_p < 0.33:287 self.transforms = CVRandomRotation(degrees=degrees)288 elif type_p < 0.66:289 self.transforms = CVRandomAffine(degrees=degrees, translate=translate, scale=scale, shear=shear)290 else:291 self.transforms = CVRandomPerspective(distortion=distortion)292 293 def __call__(self, img):294 if random.random() < self.p:295 img = np.array(img)296 return Image.fromarray(self.transforms(img))297 else: return img298 299class CVDeterioration(object):300 def __init__(self, var, degrees, factor, p=0.5):301 self.p = p302 transforms = []303 if var is not None:304 transforms.append(CVGaussianNoise(var=var))305 if degrees is not None:306 transforms.append(CVMotionBlur(degrees=degrees))307 if factor is not None:308 transforms.append(CVRescale(factor=factor))309 310 random.shuffle(transforms)311 transforms = Compose(transforms)312 self.transforms = transforms313 314 def __call__(self, img):315 if random.random() < self.p:316 img = np.array(img)317 return Image.fromarray(self.transforms(img))318 else: return img319 320 321class CVColorJitter(object):322 def __init__(self, brightness=0.5, contrast=0.5, saturation=0.5, hue=0.1, p=0.5):323 self.p = p324 self.transforms = transforms.ColorJitter(brightness=brightness, contrast=contrast, 325 saturation=saturation, hue=hue)326 327 def __call__(self, img):328 if random.random() < self.p: return self.transforms(img)329 else: return img330