图像增强方法，SSD有BBOX的情况用这个方法比较好。

参考：

https://blog.csdn.net/limiyudianzi/article/details/86497345

https://github.com/cyndi088/DataAugmentation/blob/master/augmentation.py

#!/usr/bin/python
# -*- coding: iso-8859-15 -*-
import sys
import os
import shutil
import numpy as np
import imgaug as ia
import xml.etree.ElementTree as ET

from PIL import Image
from imgaug import augmenters as iaa

ia.seed(1)

def read_xml_annotation(root, image_id):
    in_file = open(os.path.join(root, image_id))
    tree = ET.parse(in_file)
    root = tree.getroot()
    bndboxlist = []

    for object in root.findall('object'):  # 找到root节点下的所有country节点
        bndbox = object.find('bndbox')  # 子节点下节点rank的值

        xmin = int(bndbox.find('xmin').text)
        xmax = int(bndbox.find('xmax').text)
        ymin = int(bndbox.find('ymin').text)
        ymax = int(bndbox.find('ymax').text)

        bndboxlist.append([xmin, ymin, xmax, ymax])

    return bndboxlist

def change_xml_annotation(root, image_id, new_target):
    new_xmin = new_target[0]
    new_ymin = new_target[1]
    new_xmax = new_target[2]
    new_ymax = new_target[3]

    in_file = open(os.path.join(root, str(image_id) + '.xml'))  # 这里root分别由两个意思
    tree = ET.parse(in_file)
    xmlroot = tree.getroot()
    object = xmlroot.find('object')
    bndbox = object.find('bndbox')
    xmin = bndbox.find('xmin')
    xmin.text = str(new_xmin)
    ymin = bndbox.find('ymin')
    ymin.text = str(new_ymin)
    xmax = bndbox.find('xmax')
    xmax.text = str(new_xmax)
    ymax = bndbox.find('ymax')
    ymax.text = str(new_ymax)
    tree.write(os.path.join(root, str("%06d" % str(id) + '.xml')))

def change_xml_list_annotation(root, image_id, new_target, saveroot, _id):
    in_file = open(os.path.join(root, str(image_id) + '.xml'))  # 这里root分别由两个意思
    tree = ET.parse(in_file)
    elem = tree.find('filename')
    elem.text = _id + '.jpg'
    xmlroot = tree.getroot()
    index = 0

    for object in xmlroot.findall('object'):  # 找到root节点下的所有country节点
        bndbox = object.find('bndbox')  # 子节点下节点rank的值

        new_xmin = new_target[index][0]
        new_ymin = new_target[index][1]
        new_xmax = new_target[index][2]
        new_ymax = new_target[index][3]

        xmin = bndbox.find('xmin')
        xmin.text = str(new_xmin)
        ymin = bndbox.find('ymin')
        ymin.text = str(new_ymin)
        xmax = bndbox.find('xmax')
        xmax.text = str(new_xmax)
        ymax = bndbox.find('ymax')
        ymax.text = str(new_ymax)

        index = index + 1

    tree.write(os.path.join(saveroot, _id + '.xml'))

def mkdir(path):
    # 去除首位空格
    path = path.strip()
    # 去除尾部 \ 符号
    path = path.rstrip("\\")
    # 判断路径是否存在
    isExists = os.path.exists(path)
    # 判断结果
    if not isExists:
        # 如果不存在则创建目录
        os.makedirs(path)
        print(path + ' 创建成功')
        return True
    else:
        # 如果目录存在则不创建，并提示目录已存在
        print(path + ' 目录已存在')
        return False

if __name__ == "__main__":
#    reload(sys)                         # 2
#    sys.setdefaultencoding('utf-8')     # 3

    IMG_DIR = "./JPEGImages"  # 需要增强的影像文件夹路径
    XML_DIR = "./Annotations"  # 需要增强的XML文件夹路径

    AUG_IMG_DIR = "./jpegtest"  # 存储增强后的影像文件夹路径
    AUG_XML_DIR = "./annotationstest"  # 存储增强后的XML文件夹路径

    try:
        shutil.rmtree(AUG_IMG_DIR)
    except FileNotFoundError as e:
        a = 1
    mkdir(AUG_IMG_DIR)

    try:
        shutil.rmtree(AUG_XML_DIR)
    except FileNotFoundError as e:
        a = 1
    mkdir(AUG_XML_DIR)

    AUGLOOP = 8  # 每张影像增强的数量

    boxes_img_aug_list = []
    new_bndbox = []
    new_bndbox_list = []

    # 影像增强
    seq = iaa.Sequential([
        iaa.Flipud(0.5),  # vertically flip 20% of all images
        iaa.Fliplr(0.5),  # 镜像
        iaa.Multiply((1.2, 1.5)),  # change brightness, doesn't affect BBs
        iaa.GaussianBlur(sigma=(0, 3.0)),  # iaa.GaussianBlur(0.5),
        iaa.Affine(
            translate_px={"x": 15, "y": 15},
            scale=(0.8, 0.95)
            #rotate=(-5, 5)
        )  # translate by 40/60px on x/y axis, and scale to 50-70%, affects BBs
    ])

    for root, sub_folders, files in os.walk(XML_DIR):

        for name in files:

            bndbox = read_xml_annotation(XML_DIR, name)

            for epoch in range(AUGLOOP):
                seq_det = seq.to_deterministic()  # 保持坐标和图像同步改变，而不是随机
                # 读取图片
                img = Image.open(os.path.join(IMG_DIR, name[:-4] + '.jpg'))
                # sp = img.size
                img = np.asarray(img)
                # bndbox 坐标增强
                for i in range(len(bndbox)):
                    bbs = ia.BoundingBoxesOnImage([
                        ia.BoundingBox(x1=bndbox[i][0], y1=bndbox[i][1], x2=bndbox[i][2], y2=bndbox[i][3]),
                    ], shape=img.shape)

                    bbs_aug = seq_det.augment_bounding_boxes([bbs])[0]
                    boxes_img_aug_list.append(bbs_aug)

                    # new_bndbox_list:[[x1,y1,x2,y2],...[],[]]
                    n_x1 = int(max(1, min(img.shape[1], bbs_aug.bounding_boxes[0].x1)))
                    n_y1 = int(max(1, min(img.shape[0], bbs_aug.bounding_boxes[0].y1)))
                    n_x2 = int(max(1, min(img.shape[1], bbs_aug.bounding_boxes[0].x2)))
                    n_y2 = int(max(1, min(img.shape[0], bbs_aug.bounding_boxes[0].y2)))
                    if n_x1 == 1 and n_x1 == n_x2:
                        n_x2 += 1
                    if n_y1 == 1 and n_y2 == n_y1:
                        n_y2 += 1
                    if n_x1 >= n_x2 or n_y1 >= n_y2:
                        print('error', name)
                    new_bndbox_list.append([n_x1, n_y1, n_x2, n_y2])
                # 存储变化后的图片
                image_aug = seq_det.augment_images([img])[0]
                path = os.path.join(AUG_IMG_DIR, name[:-4] + '_' + str(epoch) + '.jpg')
                image_auged = bbs.draw_on_image(image_aug, thickness=0)
                Image.fromarray(image_auged).save(path)

                # 存储变化后的XML
                change_xml_list_annotation(XML_DIR, name[:-4], new_bndbox_list, AUG_XML_DIR,
                                           name[:-4] + '_' + str(epoch))
                print(name[:-4] + '_' + str(epoch) + '.jpg')
                new_bndbox_list = []

--------------------------

在图像增强序列中，把一下的一些方法用上，效果更强

seq = iaa.Sequential(
    [
        iaa.Fliplr(0.5), # 对50%的图像进行镜像翻转
        iaa.Flipud(0.2), # 对20%的图像做左右翻转

        sometimes(iaa.Crop(percent=(0, 0.1))),
       #这里沿袭我们上面提到的sometimes，对随机的一部分图像做crop操作
        # crop的幅度为0到10%

        sometimes(iaa.Affine(                          #对一部分图像做仿射变换
            scale={"x": (0.8, 1.2), "y": (0.8, 1.2)},#图像缩放为80%到120%之间
            translate_percent={"x": (-0.2, 0.2), "y": (-0.2, 0.2)}, #平移±20%之间
            rotate=(-45, 45),   #旋转±45度之间
            shear=(-16, 16),    #剪切变换±16度，（矩形变平行四边形）
            order=[0, 1],   #使用最邻近差值或者双线性差值
            cval=(0, 255),  #全白全黑填充
            mode=ia.ALL    #定义填充图像外区域的方法
        )),
        
        # 使用下面的0个到5个之间的方法去增强图像。注意SomeOf的用法
        iaa.SomeOf((0, 5),
            [
                # 将部分图像进行超像素的表示。o(╥﹏╥)o用超像素增强作者还是第一次见，比较孤陋寡闻
                sometimes(
                    iaa.Superpixels(
                        p_replace=(0, 1.0),
                        n_segments=(20, 200)
                    )
                ),

                #用高斯模糊，均值模糊，中值模糊中的一种增强。注意OneOf的用法
                iaa.OneOf([
                    iaa.GaussianBlur((0, 3.0)),
                    iaa.AverageBlur(k=(2, 7)), # 核大小2~7之间，k=((5, 7), (1, 3))时，核高度5~7，宽度1~3
                    iaa.MedianBlur(k=(3, 11)),
                ]),

                #锐化处理
                iaa.Sharpen(alpha=(0, 1.0), lightness=(0.75, 1.5)),

                #浮雕效果
                iaa.Emboss(alpha=(0, 1.0), strength=(0, 2.0)),

                #边缘检测，将检测到的赋值0或者255然后叠在原图上
                sometimes(iaa.OneOf([
                    iaa.EdgeDetect(alpha=(0, 0.7)),
                    iaa.DirectedEdgeDetect(
                        alpha=(0, 0.7), direction=(0.0, 1.0)
                    ),
                ])),

                # 加入高斯噪声
                iaa.AdditiveGaussianNoise(
                    loc=0, scale=(0.0, 0.05*255), per_channel=0.5
                ),

                # 将1%到10%的像素设置为黑色
              # 或者将3%到15%的像素用原图大小2%到5%的黑色方块覆盖
                iaa.OneOf([
                    iaa.Dropout((0.01, 0.1), per_channel=0.5),
                    iaa.CoarseDropout(
                        (0.03, 0.15), size_percent=(0.02, 0.05),
                        per_channel=0.2
                    ),
                ]),

                #5%的概率反转像素的强度，即原来的强度为v那么现在的就是255-v
                iaa.Invert(0.05, per_channel=True),

                # 每个像素随机加减-10到10之间的数
                iaa.Add((-10, 10), per_channel=0.5),

                # 像素乘上0.5或者1.5之间的数字.
                iaa.Multiply((0.5, 1.5), per_channel=0.5),

                # 将整个图像的对比度变为原来的一半或者二倍
                iaa.ContrastNormalization((0.5, 2.0), per_channel=0.5),

                # 将RGB变成灰度图然后乘alpha加在原图上
                iaa.Grayscale(alpha=(0.0, 1.0)),

                #把像素移动到周围的地方。这个方法在mnist数据集增强中有见到
                sometimes(
                    iaa.ElasticTransformation(alpha=(0.5, 3.5), sigma=0.25)
                ),

                # 扭曲图像的局部区域
                sometimes(iaa.PiecewiseAffine(scale=(0.01, 0.05)))
            ],
            
            random_order=True # 随机的顺序把这些操作用在图像上
        )
    ],
    random_order=True # 随机的顺序把这些操作用在图像上
)
=====================挑选整理后可以看如下===================

#!/usr/bin/python
# -*- coding: iso-8859-15 -*-
import sys
import os
import shutil
import numpy as np
import imgaug as ia
import xml.etree.ElementTree as ET

from PIL import Image
from imgaug import augmenters as iaa
sometimes = lambda aug: iaa.Sometimes(0.1, aug)

ia.seed(1)

def read_xml_annotation(root, image_id):
    in_file = open(os.path.join(root, image_id))
    tree = ET.parse(in_file)
    root = tree.getroot()
    bndboxlist = []

    for object in root.findall('object'):  # 找到root节点下的所有country节点
        bndbox = object.find('bndbox')  # 子节点下节点rank的值

        xmin = int(bndbox.find('xmin').text)
        xmax = int(bndbox.find('xmax').text)
        ymin = int(bndbox.find('ymin').text)
        ymax = int(bndbox.find('ymax').text)

        bndboxlist.append([xmin, ymin, xmax, ymax])

    return bndboxlist

def change_xml_annotation(root, image_id, new_target):
    new_xmin = new_target[0]
    new_ymin = new_target[1]
    new_xmax = new_target[2]
    new_ymax = new_target[3]

    in_file = open(os.path.join(root, str(image_id) + '.xml'))  # 这里root分别由两个意思
    tree = ET.parse(in_file)
    xmlroot = tree.getroot()
    object = xmlroot.find('object')
    bndbox = object.find('bndbox')
    xmin = bndbox.find('xmin')
    xmin.text = str(new_xmin)
    ymin = bndbox.find('ymin')
    ymin.text = str(new_ymin)
    xmax = bndbox.find('xmax')
    xmax.text = str(new_xmax)
    ymax = bndbox.find('ymax')
    ymax.text = str(new_ymax)
    tree.write(os.path.join(root, str("%06d" % str(id) + '.xml')))

def change_xml_list_annotation(root, image_id, new_target, saveroot, _id):
    in_file = open(os.path.join(root, str(image_id) + '.xml'))  # 这里root分别由两个意思
    tree = ET.parse(in_file)
    elem = tree.find('filename')
    elem.text = _id + '.jpg'
    xmlroot = tree.getroot()
    index = 0

    for object in xmlroot.findall('object'):  # 找到root节点下的所有country节点
        bndbox = object.find('bndbox')  # 子节点下节点rank的值

        new_xmin = new_target[index][0]
        new_ymin = new_target[index][1]
        new_xmax = new_target[index][2]
        new_ymax = new_target[index][3]

        xmin = bndbox.find('xmin')
        xmin.text = str(new_xmin)
        ymin = bndbox.find('ymin')
        ymin.text = str(new_ymin)
        xmax = bndbox.find('xmax')
        xmax.text = str(new_xmax)
        ymax = bndbox.find('ymax')
        ymax.text = str(new_ymax)

        index = index + 1

    tree.write(os.path.join(saveroot, _id + '.xml'))

def mkdir(path):
    # 去除首位空格
    path = path.strip()
    # 去除尾部 \ 符号
    path = path.rstrip("\\")
    # 判断路径是否存在
    isExists = os.path.exists(path)
    # 判断结果
    if not isExists:
        # 如果不存在则创建目录
        os.makedirs(path)
        print(path + ' 创建成功')
        return True
    else:
        # 如果目录存在则不创建，并提示目录已存在
        print(path + ' 目录已存在')
        return False

if __name__ == "__main__":
#    reload(sys)                         # 2
#    sys.setdefaultencoding('utf-8')     # 3

    IMG_DIR = "./JPEGImages"  # 需要增强的影像文件夹路径
    XML_DIR = "./Annotations"  # 需要增强的XML文件夹路径

    AUG_IMG_DIR = "./jpegtest"  # 存储增强后的影像文件夹路径
    AUG_XML_DIR = "./annotationstest"  # 存储增强后的XML文件夹路径

    try:
        shutil.rmtree(AUG_IMG_DIR)
    except FileNotFoundError as e:
        a = 1
    mkdir(AUG_IMG_DIR)

    try:
        shutil.rmtree(AUG_XML_DIR)
    except FileNotFoundError as e:
        a = 1
    mkdir(AUG_XML_DIR)

    AUGLOOP = 8  # 每张影像增强的数量

    boxes_img_aug_list = []
    new_bndbox = []
    new_bndbox_list = []

    # 影像增强
    seq = iaa.Sequential([
        iaa.Flipud(0.5),  # vertically flip 20% of all images
        iaa.Fliplr(0.5),  # 镜像
        sometimes(iaa.Multiply((1.2, 1.5))),  # change brightness, doesn't affect BBs
        sometimes(iaa.Sharpen(alpha=(0, 1.0), lightness=(0.75, 1.5))),
        sometimes(iaa.GaussianBlur(sigma=(0, 3.0))),  # iaa.GaussianBlur(0.5),
        sometimes(iaa.ContrastNormalization((0.5, 2.0), per_channel=0.5)),
        iaa.Scale(1.50,
          interpolation='cubic',
          name=None,
          deterministic=False,
          random_state=None),
        #iaa.Affine(
        #    translate_px={"x": 15, "y": 15},
        #    scale=(1.5, 1.5)
        #    #rotate=(-5, 5)
        #)  # translate by 40/60px on x/y axis, and scale to 50-70%, affects BBs
    ])

    for root, sub_folders, files in os.walk(XML_DIR):

        for name in files:

            bndbox = read_xml_annotation(XML_DIR, name)

            for epoch in range(AUGLOOP):
                seq_det = seq.to_deterministic()  # 保持坐标和图像同步改变，而不是随机
                # 读取图片
                img = Image.open(os.path.join(IMG_DIR, name[:-4] + '.jpg'))
                # sp = img.size
                img = np.asarray(img)
                # bndbox 坐标增强
                for i in range(len(bndbox)):
                    bbs = ia.BoundingBoxesOnImage([
                        ia.BoundingBox(x1=bndbox[i][0], y1=bndbox[i][1], x2=bndbox[i][2], y2=bndbox[i][3]),
                    ], shape=img.shape)

                    bbs_aug = seq_det.augment_bounding_boxes([bbs])[0]
                    boxes_img_aug_list.append(bbs_aug)

                    # new_bndbox_list:[[x1,y1,x2,y2],...[],[]]
                    n_x1 = int(max(1, min(img.shape[1], bbs_aug.bounding_boxes[0].x1)))
                    n_y1 = int(max(1, min(img.shape[0], bbs_aug.bounding_boxes[0].y1)))
                    n_x2 = int(max(1, min(img.shape[1], bbs_aug.bounding_boxes[0].x2)))
                    n_y2 = int(max(1, min(img.shape[0], bbs_aug.bounding_boxes[0].y2)))
                    if n_x1 == 1 and n_x1 == n_x2:
                        n_x2 += 1
                    if n_y1 == 1 and n_y2 == n_y1:
                        n_y2 += 1
                    if n_x1 >= n_x2 or n_y1 >= n_y2:
                        print('error', name)
                    new_bndbox_list.append([n_x1, n_y1, n_x2, n_y2])
                # 存储变化后的图片
                image_aug = seq_det.augment_images([img])[0]
                path = os.path.join(AUG_IMG_DIR, name[:-4] + '_' + str(epoch) + '.jpg')
                image_auged = bbs.draw_on_image(image_aug, thickness=0)
                Image.fromarray(image_auged).save(path)

                # 存储变化后的XML
                change_xml_list_annotation(XML_DIR, name[:-4], new_bndbox_list, AUG_XML_DIR,
                                           name[:-4] + '_' + str(epoch))
                print(name[:-4] + '_' + str(epoch) + '.jpg')
                new_bndbox_list = []