Albumentations: Fast and Flexible Image Augmentations.

Albumentations是一个为图像的数据增强设计的python库，安装如下：

pip install albumentations

1. Albumentations中的数据增强方法

Albumentations中的数据增强方法可以分为像素级的变换(pixel-level transforms)和空间级的变换(spatial-level transforms)两类。

⚪ pixel-level transforms

像素级的变换只改变图像的整体像素值，不影响图像的标签（如mask,检测框,关键点）。适用于图像分类等任务。

格式转换

引入全局噪声

平滑滤波：模糊图像

锐化滤波：增强轮廓

对比度变换

颜色变换

下游任务

⚪ spatial-level transforms

空间级的变换同时改变图像及其标注(如mask,检测框,关键点)，适用于图像分割、目标检测、姿态估计等任务。

Transform Image Masks BBoxes Keypoints
Affine ✓ ✓ ✓ ✓
CenterCrop ✓ ✓ ✓ ✓
CoarseDropout ✓ ✓    
Crop ✓ ✓ ✓ ✓
CropAndPad ✓ ✓ ✓ ✓
CropNonEmptyMaskIfExists ✓ ✓ ✓ ✓
ElasticTransform ✓ ✓    
Flip ✓ ✓ ✓ ✓
GridDistortion ✓ ✓    
GridDropout ✓ ✓    
HorizontalFlip ✓ ✓ ✓ ✓
Lambda ✓ ✓ ✓ ✓
LongestMaxSize ✓ ✓ ✓ ✓
MaskDropout ✓ ✓    
NoOp ✓ ✓ ✓ ✓
OpticalDistortion ✓ ✓    
PadIfNeeded ✓ ✓ ✓ ✓
Perspective ✓ ✓ ✓ ✓
PiecewiseAffine ✓ ✓ ✓ ✓
RandomCrop ✓ ✓ ✓ ✓
RandomCropNearBBox ✓ ✓ ✓ ✓
RandomGridShuffle ✓ ✓    
RandomResizedCrop ✓ ✓ ✓ ✓
RandomRotate90 ✓ ✓ ✓ ✓
RandomScale ✓ ✓ ✓ ✓
RandomSizedBBoxSafeCrop ✓ ✓ ✓  
RandomSizedCrop ✓ ✓ ✓ ✓
Resize ✓ ✓ ✓ ✓
Rotate ✓ ✓ ✓ ✓
SafeRotate ✓ ✓ ✓ ✓
ShiftScaleRotate ✓ ✓ ✓ ✓
SmallestMaxSize ✓ ✓ ✓ ✓
Transpose ✓ ✓ ✓ ✓
VerticalFlip ✓ ✓ ✓ ✓

2. Albumentations的使用

Albumentations的简单使用如下：

import albumentations as A
import cv2

# Declare an augmentation pipeline
transform = A.Compose([
    A.RandomCrop(width=256, height=256), # 随机裁剪
    A.HorizontalFlip(p=0.5), # 随机水平翻转
    A.RandomBrightnessContrast(p=0.2), # 随机明亮对比度
    A.OneOf([
        A.Blur(blur_limit=3, p=0.1), # 使用随机大小的内核模糊图像
        A.MedianBlur(blur_limit=3, p=0.1), # 中值滤波
    ], p=0.2),
])

# Read an image with OpenCV and convert it to the RGB colorspace
image = cv2.imread("image.jpg")
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)

# Augment an image
transformed = transform(image=image)
transformed_image = transformed["image"]

Albumentations的两个主要方法：

  1. A.Compose：顺序执行内部的变换
  2. A.OneOf：随机选择一种变换执行

Albumentations已经集成在mmdetection框架下。使用时直接修改config文件内的train_pipeline即可：

albu_train_transforms = [
    dict(type='HorizontalFlip', p=0.5),
    dict(type='OneOf', transforms=[
            dict(type='Blur', blur_limit=3, p=0.5),
            dict(type='MedianBlur', blur_limit=3, p=0.5),
        ],
        p=0.1),
]

train_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(type='LoadAnnotations', with_bbox=True),
    dict(type='Albu', transforms=albu_train_transforms),  # 数据增强
    dict(type='Resize', img_scale=(1333, 800), keep_ratio=True),
    dict(type='RandomFlip', flip_ratio=0.5),
    dict(
        type='Normalize',
        mean=[123.675, 116.28, 103.53],
        std=[58.395, 57.12, 57.375],
        to_rgb=True),
    dict(type='Pad', size_divisor=32),
    dict(type='DefaultFormatBundle'),
    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels'])
]