YOLOv5的Tricks | 【Trick12】YOLOv5使用的数据增强方法汇总

---

如有错误，恳请指出。

---

时隔两个多月重新看yolov5的代码显然开始力不从心，当时应该一鼓作气的整理完的。在专栏前面的内容一直介绍的是yolov5训练时候使用的一些技巧，这里用这篇博客最后归纳一下yolov5在数据增强上所使用的技巧。

在yolov3-spp专栏的时候，我介绍过yolov3-spp大致所使用的一些数据增强的方法：

• 数据增强——Mosaic（马赛克）

• 数据增强——随机旋转、平移、缩放、错切、hsv增强

在之前详细的介绍过代码，而在yolov5这里，其实代码是类似的，甚至函数的名字都没有变化，看过源码的朋友就可能知道了，改变的地方其实不是很多，所以这里就不再详细介绍代码的细节了，只是总结一下使用的数据增强。

0. 自定义数据集的整体架构

项目中，使用 create_dataloader 函数构建 dataloader 与 dataset，这个部分是整个算法的核心部分之一。实现数据增强的方法就是在构建 dataset 中设置的。

• create_dataloader函数

def create_dataloader(path, imgsz, batch_size, stride, single_cls=False, hyp=None, augment=False, cache=False, pad=0.0,
                      rect=False, rank=-1, workers=8, image_weights=False, quad=False, prefix=''):
    # Make sure only the first process in DDP process the dataset first, and the following others can use the cache
    with torch_distributed_zero_first(rank):
        dataset = LoadImagesAndLabels(path, imgsz, batch_size,
                                      augment=augment,  # augment images
                                      hyp=hyp,  # augmentation hyperparameters
                                      rect=rect,  # rectangular training
                                      cache_images=cache,
                                      single_cls=single_cls,
                                      stride=int(stride),
                                      pad=pad,
                                      image_weights=image_weights,
                                      prefix=prefix)

    batch_size = min(batch_size, len(dataset))

    # 这里对num_worker进行更改
    # nw = min([os.cpu_count(), batch_size if batch_size > 1 else 0, workers])  # number of workers
    nw = 0  # 可以适当提高这个参数0, 2, 4, 8, 16…

    sampler = torch.utils.data.distributed.DistributedSampler(dataset) if rank != -1 else None
    loader = torch.utils.data.DataLoader if image_weights else InfiniteDataLoader
    # Use torch.utils.data.DataLoader() if dataset.properties will update during training else InfiniteDataLoader()
    dataloader = loader(dataset,
                        batch_size=batch_size,
                        num_workers=nw,
                        sampler=sampler,
                        pin_memory=True,
                        collate_fn=LoadImagesAndLabels.collate_fn4 if quad else LoadImagesAndLabels.collate_fn)
    return dataloader, dataset

可以看见，在构建 dataloader 的时候，还对batch数据进行了设置，为其设置了批处理的 collate_fn 函数。核心重点，就是 LoadImagesAndLabels 函数，自定义了数据集的处理过程。下面详细对其分析。

• LoadImagesAndLabels函数

class LoadImagesAndLabels(Dataset):
    # YOLOv5 train_loader/val_loader, loads images and labels for training and validation
    cache_version = 0.5  # dataset labels *.cache version

    def __init__(self, path, img_size=640, batch_size=16, augment=False, hyp=None, rect=False, image_weights=False,
                 cache_images=False, single_cls=False, stride=32, pad=0.0, prefix=''):
        self.img_size = img_size
        self.augment = augment
        self.hyp = hyp
        self.image_weights = image_weights
        self.rect = False if image_weights else rect
        self.mosaic = self.augment and not self.rect  # load 4 images at a time into a mosaic (only during training)
        self.mosaic_border = [-img_size // 2, -img_size // 2]
        self.stride = stride
        self.path = path
        self.albumentations = Albumentations() if augment else None
		....
		# 下面省略了许多步骤，不过无伤大雅
		....

    def __len__(self):
        return len(self.img_files)

    # def __iter__(self):
    #     self.count = -1
    #     print('ran dataset iter')
    #     #self.shuffled_vector = np.random.permutation(self.nF) if self.augment else np.arange(self.nF)
    #     return self

    def __getitem__(self, index):
        index = self.indices[index]  # linear, shuffled, or image_weights

        hyp = self.hyp
        mosaic = self.mosaic and random.random() < hyp['mosaic']
        if mosaic:
            # Load mosaic
            img, labels = load_mosaic(self, index)  # use load_mosaic4
            # img, labels = load_mosaic9(self, index)   # use load_mosaic9
            shapes = None

            # MixUp augmentation
            if random.random() < hyp['mixup']:
                img, labels = mixup(img, labels, *load_mosaic(self, random.randint(0, self.n - 1)))
                # img, labels = mixup(img, labels, *load_mosaic9(self, random.randint(0, self.n - 1)))

        else:
            # Load image
            img, (h0, w0), (h, w) = load_image(self, index)

            # Letterbox
            shape = self.batch_shapes[self.batch[index]] if self.rect else self.img_size  # final letterboxed shape
            img, ratio, pad = letterbox(img, shape, auto=False, scaleup=self.augment)
            shapes = (h0, w0), ((h / h0, w / w0), pad)  # for COCO mAP rescaling

            labels = self.labels[index].copy()
            if labels.size:  # normalized xywh to pixel xyxy format
                labels[:, 1:] = xywhn2xyxy(labels[:, 1:], ratio[0] * w, ratio[1] * h, padw=pad[0], padh=pad[1])

            if self.augment:
                img, labels = random_perspective(img, labels,
                                                 degrees=hyp['degrees'],
                                                 translate=hyp['translate'],
                                                 scale=hyp['scale'],
                                                 shear=hyp['shear'],
                                                 perspective=hyp['perspective'])

        nl = len(labels)  # number of labels
        if nl:
            labels[:, 1:5] = xyxy2xywhn(labels[:, 1:5], w=img.shape[1], h=img.shape[0], clip=True, eps=1E-3)

        # create_dataloader to set augment
        if self.augment:
            # Albumentations
            img, labels = self.albumentations(img, labels)
            nl = len(labels)  # update after albumentations

            # HSV color-space
            # augment_hsv(img, hgain=hyp['hsv_h'], sgain=hyp['hsv_s'], vgain=hyp['hsv_v'])

            # Flip up-down
            if random.random() < hyp['flipud']:
                img = np.flipud(img)
                if nl:
                    labels[:, 2] = 1 - labels[:, 2]

            # Flip left-right
            if random.random() < hyp['fliplr']:
                img = np.fliplr(img)
                if nl:
                    labels[:, 1] = 1 - labels[:, 1]

            # Cutouts
            # labels = cutout(img, labels, p=0.5)

        labels_out = torch.zeros((nl, 6))
        if nl:
            labels_out[:, 1:] = torch.from_numpy(labels)

        # Convert
        img = img.transpose((2, 0, 1))[::-1]  # HWC to CHW, BGR to RGB
        img = np.ascontiguousarray(img)

        return torch.from_numpy(img), labels_out, self.img_files[index], shapes

    @staticmethod
    def collate_fn(batch):
        img, label, path, shapes = zip(*batch)  # transposed
        for i, l in enumerate(label):
            l[:, 0] = i  # add target image index for build_targets()
        return torch.stack(img, 0), torch.cat(label, 0), path, shapes

    @staticmethod
    def collate_fn4(batch):
        img, label, path, shapes = zip(*batch)  # transposed
        n = len(shapes) // 4
        img4, label4, path4, shapes4 = [], [], path[:n], shapes[:n]

        ho = torch.tensor([[0., 0, 0, 1, 0, 0]])
        wo = torch.tensor([[0., 0, 1, 0, 0, 0]])
        s = torch.tensor([[1, 1, .5, .5, .5, .5]])  # scale
        for i in range(n):  # zidane torch.zeros(16,3,720,1280)  # BCHW
            i *= 4
            if random.random() < 0.5:
                im = F.interpolate(img[i].unsqueeze(0).float(), scale_factor=2., mode='bilinear', align_corners=False)[
                    0].type(img[i].type())
                l = label[i]
            else:
                im = torch.cat((torch.cat((img[i], img[i + 1]), 1), torch.cat((img[i + 2], img[i + 3]), 1)), 2)
                l = torch.cat((label[i], label[i + 1] + ho, label[i + 2] + wo, label[i + 3] + ho + wo), 0) * s
            img4.append(im)
            label4.append(l)

        for i, l in enumerate(label4):
            l[:, 0] = i  # add target image index for build_targets()

        return torch.stack(img4, 0), torch.cat(label4, 0), path4, shapes4

大致的函数代码如上所示，但是其实有很多的繁琐步骤。self.rect 是都进行矩形训练，但是一般来说是进行矩形推理来加快推理速度。cache_image 是为了缓存图像，以空间换时间。这些内容我都省略掉，没有贴上来。

自定义数据集的重点是 __getitem__ 函数，各种数据增强的方式就是在这里进行的。所以我对部分代码进行了省略，只贴出了重要的部分。

---

1. Mosaic数据增强

这个部分之前已经介绍过了，不过值得一提的是，这里yolov5还额外提出了一个9图的mosaic操作，就是把之前的4个图像换成了9张图像，拼接在一起处理，图像更大了而且label也更多，训练一张这样的拼接图像等同与训练了9张小图。

• 操作示例

• Mosaic（4张）实现代码

def load_mosaic(self, index):
    # YOLOv5 4-mosaic loader. Loads 1 image + 3 random images into a 4-image mosaic
    labels4, segments4 = [], []
    s = self.img_size
    yc, xc = [int(random.uniform(-x, 2 * s + x)) for x in self.mosaic_border]  # mosaic center x, y
    indices = [index] + random.choices(self.indices, k=3)  # 3 additional image indices
    random.shuffle(indices)
    for i, index in enumerate(indices):
        # Load image
        img, _, (h, w) = load_image(self, index)

        # place img in img4
        if i == 0:  # top left
            img4 = np.full((s * 2, s * 2, img.shape[2]), 114, dtype=np.uint8)  # base image with 4 tiles
            x1a, y1a, x2a, y2a = max(xc - w, 0), max(yc - h, 0), xc, yc  # xmin, ymin, xmax, ymax (large image)
            x1b, y1b, x2b, y2b = w - (x2a - x1a), h - (y2a - y1a), w, h  # xmin, ymin, xmax, ymax (small image)
        elif i == 1:  # top right
            x1a, y1a, x2a, y2a = xc, max(yc - h, 0), min(xc + w, s * 2), yc
            x1b, y1b, x2b, y2b = 0, h - (y2a - y1a), min(w, x2a - x1a), h
        elif i == 2:  # bottom left
            x1a, y1a, x2a, y2a = max(xc - w, 0), yc, xc, min(s * 2, yc + h)
            x1b, y1b, x2b, y2b = w - (x2a - x1a), 0, w, min(y2a - y1a, h)
        elif i == 3:  # bottom right
            x1a, y1a, x2a, y2a = xc, yc, min(xc + w, s * 2), min(s * 2, yc + h)
            x1b, y1b, x2b, y2b = 0, 0, min(w, x2a - x1a), min(y2a - y1a, h)

        img4[y1a:y2a, x1a:x2a] = img[y1b:y2b, x1b:x2b]  # img4[ymin:ymax, xmin:xmax]
        padw = x1a - x1b
        padh = y1a - y1b

        # Labels
        labels, segments = self.labels[index].copy(), self.segments[index].copy()
        if labels.size:
            labels[:, 1:] = xywhn2xyxy(labels[:, 1:], w, h, padw, padh)  # normalized xywh to pixel xyxy format
            segments = [xyn2xy(x, w, h, padw, padh) for x in segments]
        labels4.append(labels)
        segments4.extend(segments)

    # Concat/clip labels
    labels4 = np.concatenate(labels4, 0)
    for x in (labels4[:, 1:], *segments4):
        np.clip(x, 0, 2 * s, out=x)  # clip when using random_perspective()
    # img4, labels4 = replicate(img4, labels4)  # replicate

    # Augment
    img4, labels4, segments4 = copy_paste(img4, labels4, segments4, p=self.hyp['copy_paste'])
    img4, labels4 = random_perspective(img4, labels4, segments4,
                                       degrees=self.hyp['degrees'],
                                       translate=self.hyp['translate'],
                                       scale=self.hyp['scale'],
                                       shear=self.hyp['shear'],
                                       perspective=self.hyp['perspective'],
                                       border=self.mosaic_border)  # border to remove

    return img4, labels4

• Mosaic（9张）实现代码

def load_mosaic9(self, index):
    # YOLOv5 9-mosaic loader. Loads 1 image + 8 random images into a 9-image mosaic
    labels9, segments9 = [], []
    s = self.img_size
    indices = [index] + random.choices(self.indices, k=8)  # 8 additional image indices
    random.shuffle(indices)
    for i, index in enumerate(indices):
        # Load image
        img, _, (h, w) = load_image(self, index)

        # place img in img9
        if i == 0:  # center
            img9 = np.full((s * 3, s * 3, img.shape[2]), 114, dtype=np.uint8)  # base image with 4 tiles
            h0, w0 = h, w
            c = s, s, s + w, s + h  # xmin, ymin, xmax, ymax (base) coordinates
        elif i == 1:  # top
            c = s, s - h, s + w, s
        elif i == 2:  # top right
            c = s + wp, s - h, s + wp + w, s
        elif i == 3:  # right
            c = s + w0, s, s + w0 + w, s + h
        elif i == 4:  # bottom right
            c = s + w0, s + hp, s + w0 + w, s + hp + h
        elif i == 5:  # bottom
            c = s + w0 - w, s + h0, s + w0, s + h0 + h
        elif i == 6:  # bottom left
            c = s + w0 - wp - w, s + h0, s + w0 - wp, s + h0 + h
        elif i == 7:  # left
            c = s - w, s + h0 - h, s, s + h0
        elif i == 8:  # top left
            c = s - w, s + h0 - hp - h, s, s + h0 - hp

        padx, pady = c[:2]
        x1, y1, x2, y2 = [max(x, 0) for x in c]  # allocate coords

        # Labels
        labels, segments = self.labels[index].copy(), self.segments[index].copy()
        if labels.size:
            labels[:, 1:] = xywhn2xyxy(labels[:, 1:], w, h, padx, pady)  # normalized xywh to pixel xyxy format
            segments = [xyn2xy(x, w, h, padx, pady) for x in segments]
        labels9.append(labels)
        segments9.extend(segments)

        # Image
        img9[y1:y2, x1:x2] = img[y1 - pady:, x1 - padx:]  # img9[ymin:ymax, xmin:xmax]
        hp, wp = h, w  # height, width previous

    # Offset
    yc, xc = [int(random.uniform(0, s)) for _ in self.mosaic_border]  # mosaic center x, y
    img9 = img9[yc:yc + 2 * s, xc:xc + 2 * s]

    # Concat/clip labels
    labels9 = np.concatenate(labels9, 0)
    labels9[:, [1, 3]] -= xc
    labels9[:, [2, 4]] -= yc
    c = np.array([xc, yc])  # centers
    segments9 = [x - c for x in segments9]

    for x in (labels9[:, 1:], *segments9):
        np.clip(x, 0, 2 * s, out=x)  # clip when using random_perspective()
    # img9, labels9 = replicate(img9, labels9)  # replicate

    # Augment
    img9, labels9 = random_perspective(img9, labels9, segments9,
                                       degrees=self.hyp['degrees'],
                                       translate=self.hyp['translate'],
                                       scale=self.hyp['scale'],
                                       shear=self.hyp['shear'],
                                       perspective=self.hyp['perspective'],
                                       border=self.mosaic_border)  # border to remove

    return img9, labels9

使用这两个方法的方式很简单，只需要改变两个地方就可以了，如下所示：

mosaic = self.mosaic and random.random() < hyp['mosaic']
if mosaic:
	# Load mosaic
	img, labels = load_mosaic(self, index)  # use load_mosaic4
	# img, labels = load_mosaic9(self, index)   # use load_mosaic9
	shapes = None
	
	# MixUp augmentation
	if random.random() < hyp['mixup']:
	    img, labels = mixup(img, labels, *load_mosaic(self, random.randint(0, self.n - 1)))
	    # img, labels = mixup(img, labels, *load_mosaic9(self, random.randint(0, self.n - 1)))

需要注意，这里的Mosaic函数并不只有Mosaic操作，还包含了仿射变换 random_perspective 与 copy_paste 操作，下面会介绍到。

---

2. Copy paste数据增强

中文名叫复制粘贴大法，将部分目标随机的粘贴到图片中，前提是数据要有segments数据才行，即每个目标的实例分割信息。下面是Copy paste原论文中的示意图。

• 操作示例

• 实现代码

def copy_paste(im, labels, segments, p=0.5):
    # Implement Copy-Paste augmentation https://arxiv.org/abs/2012.07177, labels as nx5 np.array(cls, xyxy)
    n = len(segments)
    if p and n:
        h, w, c = im.shape  # height, width, channels
        im_new = np.zeros(im.shape, np.uint8)
        for j in random.sample(range(n), k=round(p * n)):
            l, s = labels[j], segments[j]
            box = w - l[3], l[2], w - l[1], l[4]
            ioa = bbox_ioa(box, labels[:, 1:5])  # intersection over area
            if (ioa < 0.30).all():  # allow 30% obscuration of existing labels
                labels = np.concatenate((labels, [[l[0], *box]]), 0)
                segments.append(np.concatenate((w - s[:, 0:1], s[:, 1:2]), 1))
                cv2.drawContours(im_new, [segments[j].astype(np.int32)], -1, (255, 255, 255), cv2.FILLED)

        result = cv2.bitwise_and(src1=im, src2=im_new)
        result = cv2.flip(result, 1)  # augment segments (flip left-right)
        i = result > 0  # pixels to replace
        # i[:, :] = result.max(2).reshape(h, w, 1)  # act over ch
        im[i] = result[i]  # cv2.imwrite('debug.jpg', im)  # debug

    return im, labels, segments

在加载马赛克数据增强的时候，是自动使用这个方法的：

# Augment
img4, labels4, segments4 = copy_paste(img4, labels4, segments4, p=self.hyp['copy_paste'])

如果选择不用，可以直接选择注释掉即可。一般来说，在训练自定义数据集的时候，肯定没有相关的掩码，所以其实也没有用到。

---

3. Random affine仿射变换

yolov5的仿射变换包含随机旋转、平移、缩放、错切操作，和yolov3-spp一样，代码都没有改变。据配置文件里的超参数发现只使用了Scale和Translation即缩放和平移。

• 操作示例

• 实现代码

def random_perspective(im, targets=(), segments=(), degrees=10, translate=.1, scale=.1, shear=10, perspective=0.0,
                       border=(0, 0)):
    # torchvision.transforms.RandomAffine(degrees=(-10, 10), translate=(.1, .1), scale=(.9, 1.1), shear=(-10, 10))
    # targets = [cls, xyxy]

    height = im.shape[0] + border[0] * 2  # shape(h,w,c)
    width = im.shape[1] + border[1] * 2

    # Center
    C = np.eye(3)
    C[0, 2] = -im.shape[1] / 2  # x translation (pixels)
    C[1, 2] = -im.shape[0] / 2  # y translation (pixels)

    # Perspective
    P = np.eye(3)
    P[2, 0] = random.uniform(-perspective, perspective)  # x perspective (about y)
    P[2, 1] = random.uniform(-perspective, perspective)  # y perspective (about x)

    # Rotation and Scale
    R = np.eye(3)
    a = random.uniform(-degrees, degrees)
    # a += random.choice([-180, -90, 0, 90])  # add 90deg rotations to small rotations
    s = random.uniform(1 - scale, 1 + scale)
    # s = 2 ** random.uniform(-scale, scale)
    R[:2] = cv2.getRotationMatrix2D(angle=a, center=(0, 0), scale=s)

    # Shear
    S = np.eye(3)
    S[0, 1] = math.tan(random.uniform(-shear, shear) * math.pi / 180)  # x shear (deg)
    S[1, 0] = math.tan(random.uniform(-shear, shear) * math.pi / 180)  # y shear (deg)

    # Translation
    T = np.eye(3)
    T[0, 2] = random.uniform(0.5 - translate, 0.5 + translate) * width  # x translation (pixels)
    T[1, 2] = random.uniform(0.5 - translate, 0.5 + translate) * height  # y translation (pixels)

    # Combined rotation matrix
    M = T @ S @ R @ P @ C  # order of operations (right to left) is IMPORTANT
    if (border[0] != 0) or (border[1] != 0) or (M != np.eye(3)).any():  # image changed
        if perspective:
            im = cv2.warpPerspective(im, M, dsize=(width, height), borderValue=(114, 114, 114))
        else:  # affine
            im = cv2.warpAffine(im, M[:2], dsize=(width, height), borderValue=(114, 114, 114))

    # Visualize
    # import matplotlib.pyplot as plt
    # ax = plt.subplots(1, 2, figsize=(12, 6))[1].ravel()
    # ax[0].imshow(im[:, :, ::-1])  # base
    # ax[1].imshow(im2[:, :, ::-1])  # warped

    # Transform label coordinates
    n = len(targets)
    if n:
        use_segments = any(x.any() for x in segments)
        new = np.zeros((n, 4))
        if use_segments:  # warp segments
            segments = resample_segments(segments)  # upsample
            for i, segment in enumerate(segments):
                xy = np.ones((len(segment), 3))
                xy[:, :2] = segment
                xy = xy @ M.T  # transform
                xy = xy[:, :2] / xy[:, 2:3] if perspective else xy[:, :2]  # perspective rescale or affine

                # clip
                new[i] = segment2box(xy, width, height)

        else:  # warp boxes
            xy = np.ones((n * 4, 3))
            xy[:, :2] = targets[:, [1, 2, 3, 4, 1, 4, 3, 2]].reshape(n * 4, 2)  # x1y1, x2y2, x1y2, x2y1
            xy = xy @ M.T  # transform
            xy = (xy[:, :2] / xy[:, 2:3] if perspective else xy[:, :2]).reshape(n, 8)  # perspective rescale or affine

            # create new boxes
            x = xy[:, [0, 2, 4, 6]]
            y = xy[:, [1, 3, 5, 7]]
            new = np.concatenate((x.min(1), y.min(1), x.max(1), y.max(1))).reshape(4, n).T

            # clip
            new[:, [0, 2]] = new[:, [0, 2]].clip(0, width)
            new[:, [1, 3]] = new[:, [1, 3]].clip(0, height)

        # filter candidates
        i = box_candidates(box1=targets[:, 1:5].T * s, box2=new.T, area_thr=0.01 if use_segments else 0.10)
        targets = targets[i]
        targets[:, 1:5] = new[i]

    return im, targets

在加载马赛克数据增强的时候，同样是自动默认同时使用这个方法的。如果不想使用，直接注释即可。

# Augment
img4, labels4, segments4 = copy_paste(img4, labels4, segments4, p=self.hyp['copy_paste'])
img4, labels4 = random_perspective(img4, labels4, segments4,
                                   degrees=self.hyp['degrees'],
                                   translate=self.hyp['translate'],
                                   scale=self.hyp['scale'],
                                   shear=self.hyp['shear'],
                                   perspective=self.hyp['perspective'],
                                   border=self.mosaic_border)  # border to remove

而且，如果选择不适应马赛克数据增强，而是选择其他的数据增强方式，仿射变换同样会被使用到。

if mosaic:
    # Load mosaic
    img, labels = load_mosaic(self, index)  # use load_mosaic4
    ...

else:
    # Load image
    img, (h0, w0), (h, w) = load_image(self, index)
	...

    if self.augment:
        img, labels = random_perspective(img, labels,
                                         degrees=hyp['degrees'],
                                         translate=hyp['translate'],
                                         scale=hyp['scale'],
                                         shear=hyp['shear'],
                                         perspective=hyp['perspective'])

其原理，可以参考之前的文章：数据增强 | 旋转、平移、缩放、错切、HSV增强

---

4. MixUp数据增强

这个比较熟悉了，就是调整透明度两张图像叠加在一起。代码中只有较大的模型才使用到了MixUp，而且每次只有10%的概率会使用到。

• 操作示例

• 实现代码

def mixup(im, labels, im2, labels2):
    # Applies MixUp augmentation https://arxiv.org/pdf/1710.09412.pdf
    r = np.random.beta(32.0, 32.0)  # mixup ratio, alpha=beta=32.0
    im = (im * r + im2 * (1 - r)).astype(np.uint8)
    labels = np.concatenate((labels, labels2), 0)
    return im, labels

可以看见，实现只需要几行代码，比较简单的。这个方法在多数的计算机视觉模型中都有使用到。

在调用马赛克处理的时候，MixUp有一定的几率会被使用到

# MixUp augmentation
if random.random() < hyp['mixup']:
    img, labels = mixup(img, labels, *load_mosaic(self, random.randint(0, self.n - 1)))
    # img, labels = mixup(img, labels, *load_mosaic9(self, random.randint(0, self.n - 1)))

---

5. HSV随机增强图像

随机增强图像HSV在 数据增强 | 旋转、平移、缩放、错切、HSV增强 这篇文章中也有介绍到。不过在yolov5中，这里默认是注释掉不使用的。

• 操作示例

• 实现代码

def augment_hsv(im, hgain=0.5, sgain=0.5, vgain=0.5):
    # HSV color-space augmentation
    if hgain or sgain or vgain:
        r = np.random.uniform(-1, 1, 3) * [hgain, sgain, vgain] + 1  # random gains
        hue, sat, val = cv2.split(cv2.cvtColor(im, cv2.COLOR_BGR2HSV))
        dtype = im.dtype  # uint8

        x = np.arange(0, 256, dtype=r.dtype)
        lut_hue = ((x * r[0]) % 180).astype(dtype)
        lut_sat = np.clip(x * r[1], 0, 255).astype(dtype)
        lut_val = np.clip(x * r[2], 0, 255).astype(dtype)

        im_hsv = cv2.merge((cv2.LUT(hue, lut_hue), cv2.LUT(sat, lut_sat), cv2.LUT(val, lut_val)))
        cv2.cvtColor(im_hsv, cv2.COLOR_HSV2BGR, dst=im)  # no return needed

在选择进行数据增强的配置中，默认注释不使用，如下所示：

# create_dataloader to set augment
if self.augment:
    # Albumentations
    img, labels = self.albumentations(img, labels)
    nl = len(labels)  # update after albumentations

    # HSV color-space
    # augment_hsv(img, hgain=hyp['hsv_h'], sgain=hyp['hsv_s'], vgain=hyp['hsv_v'])

---

6. 随机水平翻转

这个就是如字面意思，随机上下左右的水平翻转

• 操作示例

• 实现代码

# Flip up-down
if random.random() < hyp['flipud']:
    img = np.flipud(img)
    if nl:
        labels[:, 2] = 1 - labels[:, 2]

# Flip left-right
if random.random() < hyp['fliplr']:
    img = np.fliplr(img)
    if nl:
        labels[:, 1] = 1 - labels[:, 1]

---

7. Cutout数据增强

Cutout是一种新的正则化方法。训练时随机把图片的一部分减掉，这样能提高模型的鲁棒性。它的来源是计算机视觉任务中经常遇到的物体遮挡问题。通过cutout生成一些类似被遮挡的物体，不仅可以让模型在遇到遮挡问题时表现更好，还能让模型在做决定时更多地考虑环境。

Cutout数据增强在之前也见过很多次了。在yolov5的代码中默认也是不启用的。

• 操作实例

• 实现代码

def cutout(im, labels, p=0.5):
    # Applies image cutout augmentation https://arxiv.org/abs/1708.04552
    if random.random() < p:
        h, w = im.shape[:2]
        scales = [0.5] * 1 + [0.25] * 2 + [0.125] * 4 + [0.0625] * 8 + [0.03125] * 16  # image size fraction
        for s in scales:
            mask_h = random.randint(1, int(h * s))  # create random masks
            mask_w = random.randint(1, int(w * s))

            # box
            xmin = max(0, random.randint(0, w) - mask_w // 2)
            ymin = max(0, random.randint(0, h) - mask_h // 2)
            xmax = min(w, xmin + mask_w)
            ymax = min(h, ymin + mask_h)

            # apply random color mask
            im[ymin:ymax, xmin:xmax] = [random.randint(64, 191) for _ in range(3)]

            # return unobscured labels
            if len(labels) and s > 0.03:
                box = np.array([xmin, ymin, xmax, ymax], dtype=np.float32)
                ioa = bbox_ioa(box, labels[:, 1:5])  # intersection over area
                labels = labels[ioa < 0.60]  # remove >60% obscured labels

    return labels

源码中默认是不启动的

# Cutouts
# labels = cutout(img, labels, p=0.5)

---

8. Albumentations数据增强工具包

Albumentations 数据增强工具包在之前已经介绍过，见：Yolo系列 | Yolov4v5的模型结构与正负样本匹配

• github地址：https://github.com/albumentations-team/albumentations
• docs使用文档：https://albumentations.ai/docs

其涵盖了绝大部分的数据增强方式，如下：

• yolov5代码

class Albumentations:
    # YOLOv5 Albumentations class (optional, only used if package is installed)
    def __init__(self):
        self.transform = None
        try:
            import albumentations as A
            check_version(A.__version__, '1.0.3')  # version requirement

            self.transform = A.Compose([
                A.Blur(p=0.01),
                A.MedianBlur(p=0.01),
                A.ToGray(p=0.01),
                A.CLAHE(p=0.01),
                A.RandomBrightnessContrast(p=0.0),
                A.RandomGamma(p=0.0),
                A.ImageCompression(quality_lower=75, p=0.0)],
                bbox_params=A.BboxParams(format='yolo', label_fields=['class_labels']))
            logging.info(colorstr('albumentations: ') + ', '.join(f'{x}' for x in self.transform.transforms if x.p))
        except ImportError:  # package not installed, skip
            pass
        except Exception as e:
            logging.info(colorstr('albumentations: ') + f'{e}')

    def __call__(self, im, labels, p=1.0):
        if self.transform and random.random() < p:
            new = self.transform(image=im, bboxes=labels[:, 1:], class_labels=labels[:, 0])  # transformed
            im, labels = new['image'], np.array([[c, *b] for c, b in zip(new['class_labels'], new['bboxes'])])
        return im, labels

• 自己调用代码

import albumentations as A

class Albumentations:
    # YOLOv5 Albumentations class (optional, only used if package is installed)
    def __init__(self):

        self.transform = A.Compose([
            A.Blur(p=0.15),             # 随机模糊
            A.GaussianBlur(p=0.15),     # 高斯滤波器模糊
            A.MedianBlur(p=0.15),       # 中值滤波器模糊输入图像
            A.GaussNoise(p=0.15),        # 高斯噪声应用于输入图像
            A.InvertImg(0.15),          # 通过从255中减去像素值来反转输入图像

            A.ToGray(p=0.15),           # 将输入的 RGB 图像转换为灰度
            A.CLAHE(p=0.15),            # 自适应直方图均衡
            A.ChannelShuffle(p=0.15),   # 随机重新排列输入 RGB 图像的通道
            A.ColorJitter(p=0.25),      # 随机改变图像的亮度、对比度和饱和度
            A.FancyPCA(p=0.25),         # 使用FancyPCA增强RGB图像
            A.Sharpen(p=0.15),          # 锐化输入图像并将结果与​​原始图像叠加
            A.HueSaturationValue(p=0.15),           # 随机改变输入图像的色调、饱和度和值
            A.RandomBrightnessContrast(p=0.15),     # 随机改变输入图像的亮度和对比度

            # 与random_perspective函数重复
            # A.Rotate(limit=20, p=0.35), # 随机旋转
            # A.HorizontalFlip(p=0.35),   # 水平翻转
            # A.VerticalFlip(p=0.35),     # 垂直翻转
            # A.Perspective(p=0.35),      # 透视变换

            A.ImageCompression(quality_lower=75, p=0.01)],  # 减少图像的 Jpeg、WebP 压缩

            # yolo格式的边界框坐标的格式: [x_center, y_center, width, height]
            bbox_params=A.BboxParams(format='yolo', label_fields=['class_labels']))

        logging.info(colorstr('albumentations: ') + ', '.join(f'{x}' for x in self.transform.transforms if x.p))

    def __call__(self, im, labels, p=1.0):
        if self.transform and random.random() < p:
            new = self.transform(image=im, bboxes=labels[:, 1:], class_labels=labels[:, 0])  # transformed
            im, labels = new['image'], np.array([[c, *b] for c, b in zip(new['class_labels'], new['bboxes'])])
        return im, labels

可以看见，使用方法类似于pytorch的transform。使用的方法是类似的。

不过，在Albumentations提供的数据增强方式比pytorch官方的更多，使用也比较方便。

---

参考资料：

1. YOLOv5网络详解

2. YOLOv5 (6.0/6.1) brief summary

3. Yolo系列 | Yolov4v5的模型结构与正负样本匹配

4. 数据增强 | 旋转、平移、缩放、错切、HSV增强

5. 【Trick7】数据增强——Mosaic（马赛克）

6. 【Trick8】数据增强——随机旋转、平移、缩放、错切、hsv增强