Bai360/Cotton2
0
1# YOLOv5 🚀 by Ultralytics, GPL-3.0 license2"""3Image augmentation functions4"""5 6import math7import random8 9import cv210import numpy as np11import torch12import torchvision.transforms as T13import torchvision.transforms.functional as TF14 15from utils.general import LOGGER, check_version, colorstr, resample_segments, segment2box, xywhn2xyxy16from utils.metrics import bbox_ioa17 18IMAGENET_MEAN = 0.485, 0.456, 0.406 # RGB mean19IMAGENET_STD = 0.229, 0.224, 0.225 # RGB standard deviation20 21 22class Albumentations:23 # YOLOv5 Albumentations class (optional, only used if package is installed)24 def __init__(self, size=640):25 self.transform = None26 prefix = colorstr('albumentations: ')27 try:28 import albumentations as A29 check_version(A.__version__, '1.0.3', hard=True) # version requirement30 31 T = [32 A.RandomResizedCrop(height=size, width=size, scale=(0.8, 1.0), ratio=(0.9, 1.11), p=0.0),33 A.Blur(p=0.01),34 A.MedianBlur(p=0.01),35 A.ToGray(p=0.01),36 A.CLAHE(p=0.01),37 A.RandomBrightnessContrast(p=0.0),38 A.RandomGamma(p=0.0),39 A.ImageCompression(quality_lower=75, p=0.0)] # transforms40 self.transform = A.Compose(T, bbox_params=A.BboxParams(format='yolo', label_fields=['class_labels']))41 42 LOGGER.info(prefix + ', '.join(f'{x}'.replace('always_apply=False, ', '') for x in T if x.p))43 except ImportError: # package not installed, skip44 pass45 except Exception as e:46 LOGGER.info(f'{prefix}{e}')47 48 def __call__(self, im, labels, p=1.0):49 if self.transform and random.random() < p:50 new = self.transform(image=im, bboxes=labels[:, 1:], class_labels=labels[:, 0]) # transformed51 im, labels = new['image'], np.array([[c, *b] for c, b in zip(new['class_labels'], new['bboxes'])])52 return im, labels53 54 55def normalize(x, mean=IMAGENET_MEAN, std=IMAGENET_STD, inplace=False):56 # Denormalize RGB images x per ImageNet stats in BCHW format, i.e. = (x - mean) / std57 return TF.normalize(x, mean, std, inplace=inplace)58 59 60def denormalize(x, mean=IMAGENET_MEAN, std=IMAGENET_STD):61 # Denormalize RGB images x per ImageNet stats in BCHW format, i.e. = x * std + mean62 for i in range(3):63 x[:, i] = x[:, i] * std[i] + mean[i]64 return x65 66 67def augment_hsv(im, hgain=0.5, sgain=0.5, vgain=0.5):68 # HSV color-space augmentation69 if hgain or sgain or vgain:70 r = np.random.uniform(-1, 1, 3) * [hgain, sgain, vgain] + 1 # random gains71 hue, sat, val = cv2.split(cv2.cvtColor(im, cv2.COLOR_BGR2HSV))72 dtype = im.dtype # uint873 74 x = np.arange(0, 256, dtype=r.dtype)75 lut_hue = ((x * r[0]) % 180).astype(dtype)76 lut_sat = np.clip(x * r[1], 0, 255).astype(dtype)77 lut_val = np.clip(x * r[2], 0, 255).astype(dtype)78 79 im_hsv = cv2.merge((cv2.LUT(hue, lut_hue), cv2.LUT(sat, lut_sat), cv2.LUT(val, lut_val)))80 cv2.cvtColor(im_hsv, cv2.COLOR_HSV2BGR, dst=im) # no return needed81 82 83def hist_equalize(im, clahe=True, bgr=False):84 # Equalize histogram on BGR image 'im' with im.shape(n,m,3) and range 0-25585 yuv = cv2.cvtColor(im, cv2.COLOR_BGR2YUV if bgr else cv2.COLOR_RGB2YUV)86 if clahe:87 c = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8, 8))88 yuv[:, :, 0] = c.apply(yuv[:, :, 0])89 else:90 yuv[:, :, 0] = cv2.equalizeHist(yuv[:, :, 0]) # equalize Y channel histogram91 return cv2.cvtColor(yuv, cv2.COLOR_YUV2BGR if bgr else cv2.COLOR_YUV2RGB) # convert YUV image to RGB92 93 94def replicate(im, labels):95 # Replicate labels96 h, w = im.shape[:2]97 boxes = labels[:, 1:].astype(int)98 x1, y1, x2, y2 = boxes.T99 s = ((x2 - x1) + (y2 - y1)) / 2 # side length (pixels)100 for i in s.argsort()[:round(s.size * 0.5)]: # smallest indices101 x1b, y1b, x2b, y2b = boxes[i]102 bh, bw = y2b - y1b, x2b - x1b103 yc, xc = int(random.uniform(0, h - bh)), int(random.uniform(0, w - bw)) # offset x, y104 x1a, y1a, x2a, y2a = [xc, yc, xc + bw, yc + bh]105 im[y1a:y2a, x1a:x2a] = im[y1b:y2b, x1b:x2b] # im4[ymin:ymax, xmin:xmax]106 labels = np.append(labels, [[labels[i, 0], x1a, y1a, x2a, y2a]], axis=0)107 108 return im, labels109 110 111def letterbox(im, new_shape=(640, 640), color=(114, 114, 114), auto=True, scaleFill=False, scaleup=True, stride=32):112 # Resize and pad image while meeting stride-multiple constraints113 shape = im.shape[:2] # current shape [height, width]114 if isinstance(new_shape, int):115 new_shape = (new_shape, new_shape)116 117 # Scale ratio (new / old)118 r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])119 if not scaleup: # only scale down, do not scale up (for better val mAP)120 r = min(r, 1.0)121 122 # Compute padding123 ratio = r, r # width, height ratios124 new_unpad = int(round(shape[1] * r)), int(round(shape[0] * r))125 dw, dh = new_shape[1] - new_unpad[0], new_shape[0] - new_unpad[1] # wh padding126 if auto: # minimum rectangle127 dw, dh = np.mod(dw, stride), np.mod(dh, stride) # wh padding128 elif scaleFill: # stretch129 dw, dh = 0.0, 0.0130 new_unpad = (new_shape[1], new_shape[0])131 ratio = new_shape[1] / shape[1], new_shape[0] / shape[0] # width, height ratios132 133 dw /= 2 # divide padding into 2 sides134 dh /= 2135 136 if shape[::-1] != new_unpad: # resize137 im = cv2.resize(im, new_unpad, interpolation=cv2.INTER_LINEAR)138 top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1))139 left, right = int(round(dw - 0.1)), int(round(dw + 0.1))140 im = cv2.copyMakeBorder(im, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color) # add border141 return im, ratio, (dw, dh)142 143 144def random_perspective(im,145 targets=(),146 segments=(),147 degrees=10,148 translate=.1,149 scale=.1,150 shear=10,151 perspective=0.0,152 border=(0, 0)):153 # torchvision.transforms.RandomAffine(degrees=(-10, 10), translate=(0.1, 0.1), scale=(0.9, 1.1), shear=(-10, 10))154 # targets = [cls, xyxy]155 156 height = im.shape[0] + border[0] * 2 # shape(h,w,c)157 width = im.shape[1] + border[1] * 2158 159 # Center160 C = np.eye(3)161 C[0, 2] = -im.shape[1] / 2 # x translation (pixels)162 C[1, 2] = -im.shape[0] / 2 # y translation (pixels)163 164 # Perspective165 P = np.eye(3)166 P[2, 0] = random.uniform(-perspective, perspective) # x perspective (about y)167 P[2, 1] = random.uniform(-perspective, perspective) # y perspective (about x)168 169 # Rotation and Scale170 R = np.eye(3)171 a = random.uniform(-degrees, degrees)172 # a += random.choice([-180, -90, 0, 90]) # add 90deg rotations to small rotations173 s = random.uniform(1 - scale, 1 + scale)174 # s = 2 ** random.uniform(-scale, scale)175 R[:2] = cv2.getRotationMatrix2D(angle=a, center=(0, 0), scale=s)176 177 # Shear178 S = np.eye(3)179 S[0, 1] = math.tan(random.uniform(-shear, shear) * math.pi / 180) # x shear (deg)180 S[1, 0] = math.tan(random.uniform(-shear, shear) * math.pi / 180) # y shear (deg)181 182 # Translation183 T = np.eye(3)184 T[0, 2] = random.uniform(0.5 - translate, 0.5 + translate) * width # x translation (pixels)185 T[1, 2] = random.uniform(0.5 - translate, 0.5 + translate) * height # y translation (pixels)186 187 # Combined rotation matrix188 M = T @ S @ R @ P @ C # order of operations (right to left) is IMPORTANT189 if (border[0] != 0) or (border[1] != 0) or (M != np.eye(3)).any(): # image changed190 if perspective:191 im = cv2.warpPerspective(im, M, dsize=(width, height), borderValue=(114, 114, 114))192 else: # affine193 im = cv2.warpAffine(im, M[:2], dsize=(width, height), borderValue=(114, 114, 114))194 195 # Visualize196 # import matplotlib.pyplot as plt197 # ax = plt.subplots(1, 2, figsize=(12, 6))[1].ravel()198 # ax[0].imshow(im[:, :, ::-1]) # base199 # ax[1].imshow(im2[:, :, ::-1]) # warped200 201 # Transform label coordinates202 n = len(targets)203 if n:204 use_segments = any(x.any() for x in segments)205 new = np.zeros((n, 4))206 if use_segments: # warp segments207 segments = resample_segments(segments) # upsample208 for i, segment in enumerate(segments):209 xy = np.ones((len(segment), 3))210 xy[:, :2] = segment211 xy = xy @ M.T # transform212 xy = xy[:, :2] / xy[:, 2:3] if perspective else xy[:, :2] # perspective rescale or affine213 214 # clip215 new[i] = segment2box(xy, width, height)216 217 else: # warp boxes218 xy = np.ones((n * 4, 3))219 xy[:, :2] = targets[:, [1, 2, 3, 4, 1, 4, 3, 2]].reshape(n * 4, 2) # x1y1, x2y2, x1y2, x2y1220 xy = xy @ M.T # transform221 xy = (xy[:, :2] / xy[:, 2:3] if perspective else xy[:, :2]).reshape(n, 8) # perspective rescale or affine222 223 # create new boxes224 x = xy[:, [0, 2, 4, 6]]225 y = xy[:, [1, 3, 5, 7]]226 new = np.concatenate((x.min(1), y.min(1), x.max(1), y.max(1))).reshape(4, n).T227 228 # clip229 new[:, [0, 2]] = new[:, [0, 2]].clip(0, width)230 new[:, [1, 3]] = new[:, [1, 3]].clip(0, height)231 232 # filter candidates233 i = box_candidates(box1=targets[:, 1:5].T * s, box2=new.T, area_thr=0.01 if use_segments else 0.10)234 targets = targets[i]235 targets[:, 1:5] = new[i]236 237 return im, targets238 239 240def copy_paste(im, labels, segments, p=0.5):241 # Implement Copy-Paste augmentation https://arxiv.org/abs/2012.07177, labels as nx5 np.array(cls, xyxy)242 n = len(segments)243 if p and n:244 h, w, c = im.shape # height, width, channels245 im_new = np.zeros(im.shape, np.uint8)246 for j in random.sample(range(n), k=round(p * n)):247 l, s = labels[j], segments[j]248 box = w - l[3], l[2], w - l[1], l[4]249 ioa = bbox_ioa(box, labels[:, 1:5]) # intersection over area250 if (ioa < 0.30).all(): # allow 30% obscuration of existing labels251 labels = np.concatenate((labels, [[l[0], *box]]), 0)252 segments.append(np.concatenate((w - s[:, 0:1], s[:, 1:2]), 1))253 cv2.drawContours(im_new, [segments[j].astype(np.int32)], -1, (1, 1, 1), cv2.FILLED)254 255 result = cv2.flip(im, 1) # augment segments (flip left-right)256 i = cv2.flip(im_new, 1).astype(bool)257 im[i] = result[i] # cv2.imwrite('debug.jpg', im) # debug258 259 return im, labels, segments260 261 262def cutout(im, labels, p=0.5):263 # Applies image cutout augmentation https://arxiv.org/abs/1708.04552264 if random.random() < p:265 h, w = im.shape[:2]266 scales = [0.5] * 1 + [0.25] * 2 + [0.125] * 4 + [0.0625] * 8 + [0.03125] * 16 # image size fraction267 for s in scales:268 mask_h = random.randint(1, int(h * s)) # create random masks269 mask_w = random.randint(1, int(w * s))270 271 # box272 xmin = max(0, random.randint(0, w) - mask_w // 2)273 ymin = max(0, random.randint(0, h) - mask_h // 2)274 xmax = min(w, xmin + mask_w)275 ymax = min(h, ymin + mask_h)276 277 # apply random color mask278 im[ymin:ymax, xmin:xmax] = [random.randint(64, 191) for _ in range(3)]279 280 # return unobscured labels281 if len(labels) and s > 0.03:282 box = np.array([xmin, ymin, xmax, ymax], dtype=np.float32)283 ioa = bbox_ioa(box, xywhn2xyxy(labels[:, 1:5], w, h)) # intersection over area284 labels = labels[ioa < 0.60] # remove >60% obscured labels285 286 return labels287 288 289def mixup(im, labels, im2, labels2):290 # Applies MixUp augmentation https://arxiv.org/pdf/1710.09412.pdf291 r = np.random.beta(32.0, 32.0) # mixup ratio, alpha=beta=32.0292 im = (im * r + im2 * (1 - r)).astype(np.uint8)293 labels = np.concatenate((labels, labels2), 0)294 return im, labels295 296 297def box_candidates(box1, box2, wh_thr=2, ar_thr=100, area_thr=0.1, eps=1e-16): # box1(4,n), box2(4,n)298 # Compute candidate boxes: box1 before augment, box2 after augment, wh_thr (pixels), aspect_ratio_thr, area_ratio299 w1, h1 = box1[2] - box1[0], box1[3] - box1[1]300 w2, h2 = box2[2] - box2[0], box2[3] - box2[1]301 ar = np.maximum(w2 / (h2 + eps), h2 / (w2 + eps)) # aspect ratio302 return (w2 > wh_thr) & (h2 > wh_thr) & (w2 * h2 / (w1 * h1 + eps) > area_thr) & (ar < ar_thr) # candidates303 304 305def classify_albumentations(306 augment=True,307 size=224,308 scale=(0.08, 1.0),309 ratio=(0.75, 1.0 / 0.75), # 0.75, 1.33310 hflip=0.5,311 vflip=0.0,312 jitter=0.4,313 mean=IMAGENET_MEAN,314 std=IMAGENET_STD,315 auto_aug=False):316 # YOLOv5 classification Albumentations (optional, only used if package is installed)317 prefix = colorstr('albumentations: ')318 try:319 import albumentations as A320 from albumentations.pytorch import ToTensorV2321 check_version(A.__version__, '1.0.3', hard=True) # version requirement322 if augment: # Resize and crop323 T = [A.RandomResizedCrop(height=size, width=size, scale=scale, ratio=ratio)]324 if auto_aug:325 # TODO: implement AugMix, AutoAug & RandAug in albumentation326 LOGGER.info(f'{prefix}auto augmentations are currently not supported')327 else:328 if hflip > 0:329 T += [A.HorizontalFlip(p=hflip)]330 if vflip > 0:331 T += [A.VerticalFlip(p=vflip)]332 if jitter > 0:333 color_jitter = (float(jitter),) * 3 # repeat value for brightness, contrast, satuaration, 0 hue334 T += [A.ColorJitter(*color_jitter, 0)]335 else: # Use fixed crop for eval set (reproducibility)336 T = [A.SmallestMaxSize(max_size=size), A.CenterCrop(height=size, width=size)]337 T += [A.Normalize(mean=mean, std=std), ToTensorV2()] # Normalize and convert to Tensor338 LOGGER.info(prefix + ', '.join(f'{x}'.replace('always_apply=False, ', '') for x in T if x.p))339 return A.Compose(T)340 341 except ImportError: # package not installed, skip342 LOGGER.warning(f'{prefix}⚠️ not found, install with `pip install albumentations` (recommended)')343 except Exception as e:344 LOGGER.info(f'{prefix}{e}')345 346 347def classify_transforms(size=224):348 # Transforms to apply if albumentations not installed349 assert isinstance(size, int), f'ERROR: classify_transforms size {size} must be integer, not (list, tuple)'350 # T.Compose([T.ToTensor(), T.Resize(size), T.CenterCrop(size), T.Normalize(IMAGENET_MEAN, IMAGENET_STD)])351 return T.Compose([CenterCrop(size), ToTensor(), T.Normalize(IMAGENET_MEAN, IMAGENET_STD)])352 353 354class LetterBox:355 # YOLOv5 LetterBox class for image preprocessing, i.e. T.Compose([LetterBox(size), ToTensor()])356 def __init__(self, size=(640, 640), auto=False, stride=32):357 super().__init__()358 self.h, self.w = (size, size) if isinstance(size, int) else size359 self.auto = auto # pass max size integer, automatically solve for short side using stride360 self.stride = stride # used with auto361 362 def __call__(self, im): # im = np.array HWC363 imh, imw = im.shape[:2]364 r = min(self.h / imh, self.w / imw) # ratio of new/old365 h, w = round(imh * r), round(imw * r) # resized image366 hs, ws = (math.ceil(x / self.stride) * self.stride for x in (h, w)) if self.auto else self.h, self.w367 top, left = round((hs - h) / 2 - 0.1), round((ws - w) / 2 - 0.1)368 im_out = np.full((self.h, self.w, 3), 114, dtype=im.dtype)369 im_out[top:top + h, left:left + w] = cv2.resize(im, (w, h), interpolation=cv2.INTER_LINEAR)370 return im_out371 372 373class CenterCrop:374 # YOLOv5 CenterCrop class for image preprocessing, i.e. T.Compose([CenterCrop(size), ToTensor()])375 def __init__(self, size=640):376 super().__init__()377 self.h, self.w = (size, size) if isinstance(size, int) else size378 379 def __call__(self, im): # im = np.array HWC380 imh, imw = im.shape[:2]381 m = min(imh, imw) # min dimension382 top, left = (imh - m) // 2, (imw - m) // 2383 return cv2.resize(im[top:top + m, left:left + m], (self.w, self.h), interpolation=cv2.INTER_LINEAR)384 385 386class ToTensor:387 # YOLOv5 ToTensor class for image preprocessing, i.e. T.Compose([LetterBox(size), ToTensor()])388 def __init__(self, half=False):389 super().__init__()390 self.half = half391 392 def __call__(self, im): # im = np.array HWC in BGR order393 im = np.ascontiguousarray(im.transpose((2, 0, 1))[::-1]) # HWC to CHW -> BGR to RGB -> contiguous394 im = torch.from_numpy(im) # to torch395 im = im.half() if self.half else im.float() # uint8 to fp16/32396 im /= 255.0 # 0-255 to 0.0-1.0397 return im398