数据增强 Data Augmentation

数据增强是一种技术，通过操作原始数据来增加输入数据。例如，对于图像，我们可以执行以下操作：旋转图像、翻转图像、添加高斯模糊等。
在深度学习项目中，寻找数据花费了相当多的时间。但在很多实际的项目中，我们难以找到充足的数据来完成任务。
为了要保证完美地完成项目，有两件事情需要做好：

1. 寻找更多的数据；
2. 数据增强。

还有一点要提一下，每一个epoch训练的数据分布都会有变化，这样可以训练模型的泛化能力。
数据增强的种类：

1. 空间几何变换类：翻转，剪切，缩放变形，仿射
2. 像素颜色变换类：噪声类、模糊类、HSV对比度变换、RGB颜色扰动、随机擦除法、超像素法（Superpixels）
3. 多样本合成类：Smote

pixel augmentation
deformation augmentation
ScaleAug
CutOut

data-augmentation

from imgaug import augmenters as iaa
# pixel augmentation
class ImageAug(object):
    def __call__(self, sample):
        image, mask = sample
        if np.random.uniform(0,1) > 0.5:
            seq = iaa.Sequential([iaa.OneOf([
                iaa.AdditiveGaussianNoise(scale=(0, 0.2 * 255)),
                iaa.Sharpen(alpha=(0.1, 0.3), lightness=(0.7, 1.3)),
                iaa.GaussianBlur(sigma=(0, 1.0))])])
            image = seq.augment_image(image)
        return image, mask

# deformation augmentation
class DeformAug(object):
    def __call__(self, sample):
        image, mask = sample
        seq = iaa.Sequential([iaa.CropAndPad(percent=(-0.05, 0.1))])
        seg_to = seq.to_deterministic()
        image = seg_to.augment_image(image)
        mask = seg_to.augment_image(mask)
        return image, mask

class ScaleAug(object):
    def __call__(self, sample):
        image, mask = sample
        scale = random.uniform(0.7, 1.5)
        h, w, _ = image.shape
        aug_image = image.copy()
        aug_mask = mask.copy()
        aug_image = cv2.resize(aug_image, (int (scale * w), int (scale * h)))
        aug_mask = cv2.resize(aug_mask, (int (scale * w), int (scale * h)))
        if (scale < 1.0):
            new_h, new_w, _ = aug_image.shape
            pre_h_pad = int((h - new_h) / 2)
            pre_w_pad = int((w - new_w) / 2)
            pad_list = [[pre_h_pad, h - new_h - pre_h_pad], [pre_w_pad, w - new_w - pre_w_pad], [0, 0]]
            aug_image = np.pad(aug_image, pad_list, mode="constant")
            aug_mask = np.pad(aug_mask, pad_list[:2], mode="constant")
        if (scale > 1.0):
            new_h, new_w, _ = aug_image.shape
            pre_h_crop = int ((new_h - h) / 2)
            pre_w_crop = int ((new_w - w) / 2)
            post_h_crop = h + pre_h_crop
            post_w_crop = w + pre_w_crop
            aug_image = aug_image[pre_h_crop:post_h_crop, pre_w_crop:post_w_crop]
            aug_mask = aug_mask[pre_h_crop:post_h_crop, pre_w_crop:post_w_crop]
        return aug_image, aug_mask


class CutOut(object):
    def __init__(self, mask_size, p):
        self.mask_size = mask_size
        self.p = p

    def __call__(self, sample):
        image, mask = sample
        mask_size_half = self.mask_size // 2
        offset = 1 if self.mask_size % 2 == 0 else 0

        h, w = image.shape[:2]
        cxmin, cxmax = mask_size_half, w + offset - mask_size_half
        cymin, cymax = mask_size_half, h + offset - mask_size_half
        cx = np.random.randint(cxmin, cxmax)
        cy = np.random.randint(cymin, cymax)
        xmin, ymin = cx - mask_size_half, cy - mask_size_half
        xmax, ymax = xmin + self.mask_size, ymin + self.mask_size
        xmin, ymin, xmax, ymax = max(0, xmin), max(0, ymin), min(w, xmax), min(h, ymax)
        if np.random.uniform(0, 1) < self.p:
            image[ymin:ymax, xmin:xmax] = (0, 0, 0)
        return image, mask


class ToTensor(object):
    def __call__(self, sample):

        image, mask = sample
        image = np.transpose(image,(2,0,1))
        image = image.astype(np.float32)
        mask = mask.astype(np.long)
        return {'image': torch.from_numpy(image.copy()),
                'label': torch.from_numpy(mask.copy())}

参考资料：
https://zhuanlan.zhihu.com/p/41679153