【目标检测实战学习】数据增强的几种方法：cutout，mixup，mosaic，rotate，HSV，随机抖动实战
最近在学习数据增强方面的东西，简单做个记录
首先需要强调的是，数据增强是目标检测流程中的一个过程，通常是在对数据集完成打标签之后，在划分数据集之前，为了增大数据集的数量，获取更多的特征，采用的一种方式。所以，在实战的过程中，不仅仅要对图像进行操作，还要对已经打好的标签（VOC数据集的xml文件）进行同样的对应操作
随机抖动，mosaic，mixup三种方法参考的是GitHub上大佬的代码，链接如下：
bubbliiiing/object-detection-augmentation: 这里面存放了一些目标检测算法的数据增强方法。如mosaic、mixup。 (github.com)
rotate，HSV，的参考
DataAugmentation_ForObjectDetect/rotated.py at master · DLLXW/DataAugmentation_ForObjectDetect · GitHub
cutout方法是看了两篇博主的博客，然后缝合出来的，链接如下：
(24条消息) 数据增强实测之cutout_一个菜鸟的奋斗的博客-CSDN博客_cutout数据增强
(24条消息) Cutout一种新的正则化方法_点PY的博客-CSDN博客_cutout方法
下面开始正题
首先需要准备好需要数据加强的原图（我使用的是jpeg格式）以及与原图一一对应的标签文件，命名必须一样
如何使用labelimg制作VOC数据集可以看我之前的贴子
【目标检测实战学习】从零开始制作并训练自己的VOC数据集，并使用Retinanet进行目标检测_Bill～QAQ～的博客-CSDN博客
1.旋转（rotate）
就是将图片进行一定角度的旋转，当旋转角度为180度或者360度时，图像的大小不会发生变化，否则图像就会呈现内切矩形的大小：
原图
旋转200度后
代码如下：
import cv2
import math
import numpy as np
import os
import glob
import json
import shutil
import xml.etree.ElementTree as ET
from xml.etree.ElementTree import ElementTree, Element
def getRotatedImg(Pi_angle,img_path,img_write_path):
img = cv2.imread(img_path)
rows, cols = img.shape[:2]
a, b = cols / 2, rows / 2
M = cv2.getRotationMatrix2D((a, b), angle, 1)
rotated_img = cv2.warpAffine(img, M, (cols, rows)) # 旋转后的图像保持大小不变
cv2.imwrite(img_write_path,rotated_img)
return a,b
def getRotatedAnno(Pi_angle,a,b,anno_path,anno_write_path):
tree = ET.parse(anno_path)
root = tree.getroot()
objects = root.findall("object")
for obj in objects:
bbox = obj.find('bndbox')
x1 = float(bbox.find('xmin').text) - 1
y1 = float(bbox.find('ymin').text) - 1
x2 = float(bbox.find('xmax').text) - 1
y2 = float(bbox.find('ymax').text) - 1
x3=x1
y3=y2
x4=x2
y4=y1
X1 = (x1 - a) * math.cos(Pi_angle) - (y1 - b) * math.sin(Pi_angle) + a
Y1 = (x1 - a) * math.sin(Pi_angle) + (y1 - b) * math.cos(Pi_angle) + b
X2 = (x2 - a) * math.cos(Pi_angle) - (y2 - b) * math.sin(Pi_angle) + a
Y2 = (x2 - a) * math.sin(Pi_angle) + (y2 - b) * math.cos(Pi_angle) + b
X3 = (x3 - a) * math.cos(Pi_angle) - (y3 - b) * math.sin(Pi_angle) + a
Y3 = (x3 - a) * math.sin(Pi_angle) + (y3 - b) * math.cos(Pi_angle) + b
X4 = (x4 - a) * math.cos(Pi_angle) - (y4 - b) * math.sin(Pi_angle) + a
Y4 = (x4 - a) * math.sin(Pi_angle) + (y4 - b) * math.cos(Pi_angle) + b
X_MIN=min(X1,X2,X3,X4)
X_MAX = max(X1, X2, X3, X4)
Y_MIN = min(Y1, Y2, Y3, Y4)
Y_MAX = max(Y1, Y2, Y3, Y4)
bbox.find('xmin').text=str(int(X_MIN))
bbox.find('ymin').text=str(int(Y_MIN))
bbox.find('xmax').text=str(int(X_MAX))
bbox.find('ymax').text=str(int(Y_MAX))
tree.write(anno_write_path) # 保存修改后的XML文件
def rotate(angle,img_dir,anno_dir,img_write_dir,anno_write_dir):
if not os.path.exists(img_write_dir):
os.makedirs(img_write_dir)
if not os.path.exists(anno_write_dir):
os.makedirs(anno_write_dir)
Pi_angle = -angle * math.pi / 180.0 # 弧度制，后面旋转坐标需要用到，注意负号！！！
img_names=os.listdir(img_dir)
for img_name in img_names:
img_path=os.path.join(img_dir,img_name)
img_write_path=os.path.join(img_write_dir,img_name[:-4]+'.jpg')
#
anno_path=os.path.join(anno_dir,img_name[:-4]+'.xml')
anno_write_path = os.path.join(anno_write_dir, img_name[:-4]+'.xml')
#
a,b=getRotatedImg(Pi_angle,img_path,img_write_path)
getRotatedAnno(Pi_angle,a,b,anno_path,anno_write_path)
angle=200
img_dir='VOCdevkit_Origin/VOC2007/JPEGImages'
anno_dir='VOCdevkit_Origin/VOC2007/Annotations'
img_write_dir='H:\object-detection-augmentation\VOCdevkit\VOC2007\JPEGImages'
anno_write_dir='H:\object-detection-augmentation\VOCdevkit\VOC2007\Annotations'
rotate(angle,img_dir,anno_dir,img_write_dir,anno_write_dir)2.Cutout
cutout就是在图像中随机选取一个或多个区域，将其裁剪掉（像素设为0或者其他）
原图：
cutout之后：
需要注意的是，因为cutout方法不会改变xml标签中bbox的位置信息，所以该方法不需要对标签文件xml进行修改
操作代码如下：
import cv2
from torchvision import transforms
import os.path
import glob
import torch
import numpy as np
class Cutout(object):
def __init__(self, n_holes=30, length=25):
self.n_holes = n_holes
self.length = length
def __call__(self, img):
h = img.size(1)
w = img.size(2)
mask = np.ones((h, w), np.float32)
for n in range(self.n_holes):
y = np.random.randint(h)
x = np.random.randint(w)
y1 = np.clip(y - self.length // 2, 0, h)
y2 = np.clip(y + self.length // 2, 0, h)
x1 = np.clip(x - self.length // 2, 0, w)
x2 = np.clip(x + self.length // 2, 0, w)
mask[y1: y2, x1: x2] = 0.
mask = torch.from_numpy(mask)
mask = mask.expand_as(img)
img = img * mask
return img
def cutout_img(inputfile,outputfile):
src=cv2.imread(inputfile,cv2.IMREAD_UNCHANGED)
src=transforms.ToTensor()(src)
#每张图片中cutout的大小
cut=Cutout(length=25,n_holes=50)
src=cut(src)
#写入
src=src.mul(255).byte()
src=src.numpy().transpose((1,2,0))
cv2.imwrite(os.path.join(outputfile,os.path.basename(inputfile)),src)
#需要进行预处理的图片路径
for inputfile in glob.glob(r'H:\object-detection-augmentation\VOCdevkit_Origin\VOC2007\JPEGImages\*jpeg'):
#保存的路径,这里设置为原文件的路径，直接覆盖
cutout_img(inputfile,r'H:\object-detection-augmentation\VOCdevkit\VOC2007\JPEGImages')
讲几个部分的参数
def __init__(self, n_holes=30, length=25):n_holes为需要生成多少个切割方块
length为每一个切割方块的大小（length*length）
mask[y1: y2, x1: x2] = 0.0.就是代表所有的切割方块都是黑色，调成其他值就会有其他的颜色（155就会呈现出类似马赛克的颜色）
在最后的路径中
#需要进行预处理的图片路径
for inputfile in glob.glob(r'H:\object-detection-augmentation\VOCdevkit_Origin\VOC2007\JPEGImages\*jpg'):
#保存的路径,这里设置为原文件的路径，直接覆盖
cutout_img(inputfile,r'H:\object-detection-augmentation\VOCdevkit\VOC2007\JPEGImages')glob.glob代表一个一个遍历这个文件夹下的图片，最后这个路径要加上\*jpg，代表读取JPEGImages下所有结尾是jpeg的文件
3.mixup
mixup就是图像融合，在数据集中随机选取两张图片，进行图像融合，同时带着他们原始图像的标签位置变动
代码如下：
import os
from random import sample
import numpy as np
from PIL import Image, ImageDraw
from utils.random_data import get_random_data, get_random_data_with_MixUp
from utils.utils import convert_annotation, get_classes
#-----------------------------------------------------------------------------------#
# Origin_VOCdevkit_path 原始标签所在的路径
# Out_VOCdevkit_path 输出标签所在的路径
# 处理后的标签为灰度图，如果设置的值太小会看不见具体情况。
#-----------------------------------------------------------------------------------#
Origin_VOCdevkit_path = "VOCdevkit_Origin"
Out_VOCdevkit_path = "VOCdevkit"
#-----------------------------------------------------------------------------------#
# Out_Num 利用mixup生成多少组图片
# input_shape 生成的图片大小
#-----------------------------------------------------------------------------------#
Out_Num = 50
input_shape = [640, 640]
#-----------------------------------------------------------------------------------#
# 下面定义了xml里面的组成模块，无需改动。
#-----------------------------------------------------------------------------------#
headstr = """\
VOC
%s
My Database
COCO
flickr
NULL
NULL
company
%d
%d
%d
0
"""
objstr = """\
"""
tailstr = '''\
'''
if __name__ == "__main__":
Origin_JPEGImages_path = os.path.join(Origin_VOCdevkit_path, "VOC2007/JPEGImages")
Origin_Annotations_path = os.path.join(Origin_VOCdevkit_path, "VOC2007/Annotations")
Out_JPEGImages_path = os.path.join(Out_VOCdevkit_path, "VOC2007/JPEGImages")
Out_Annotations_path = os.path.join(Out_VOCdevkit_path, "VOC2007/Annotations")
if not os.path.exists(Out_JPEGImages_path):
os.makedirs(Out_JPEGImages_path)
if not os.path.exists(Out_Annotations_path):
os.makedirs(Out_Annotations_path)
#---------------------------#
# 遍历标签并赋值
#---------------------------#
xml_names = os.listdir(Origin_Annotations_path)
def write_xml(anno_path, jpg_pth, head, input_shape, boxes, unique_labels, tail):
f = open(anno_path, "w")
f.write(head%(jpg_pth, input_shape[0], input_shape[1], 3))
for i, box in enumerate(boxes):
f.write(objstr%(str(unique_labels[int(box[4])]), box[0], box[1], box[2], box[3]))
f.write(tail)
#########以上部分无需改动###########################################################################################
#------------------------------#
# 循环生成xml和jpg
#------------------------------#
for index in range(Out_Num):
#------------------------------#
# 获取两个图像与标签
#------------------------------#
sample_xmls = sample(xml_names, 2)
unique_labels = get_classes(sample_xmls, Origin_Annotations_path)
jpg_name_1 = os.path.join(Origin_JPEGImages_path, os.path.splitext(sample_xmls[0])[0] + '.jpg')
jpg_name_2 = os.path.join(Origin_JPEGImages_path, os.path.splitext(sample_xmls[1])[0] + '.jpg')
xml_name_1 = os.path.join(Origin_Annotations_path, sample_xmls[0])
xml_name_2 = os.path.join(Origin_Annotations_path, sample_xmls[1])
line_1 = convert_annotation(jpg_name_1, xml_name_1, unique_labels)
line_2 = convert_annotation(jpg_name_2, xml_name_2, unique_labels)
#------------------------------#
# 各自数据增强
#------------------------------#
image_1, box_1 = get_random_data(line_1, input_shape)
image_2, box_2 = get_random_data(line_2, input_shape)
#------------------------------#
# 合并mixup
#------------------------------#
image_data, box_data = get_random_data_with_MixUp(image_1, box_1, image_2, box_2)
img = Image.fromarray(image_data.astype(np.uint8))
img.save(os.path.join(Out_JPEGImages_path, str(index) + '.jpg'))
write_xml(os.path.join(Out_Annotations_path, str(index) + '.xml'), os.path.join(Out_JPEGImages_path, str(index) + '.jpg'), \
headstr, input_shape, box_data, unique_labels, tailstr)
可以注释修改的部分都已经注释在里面了
4.mosaic
mosaic有点类似于拼图，在数据集中随机选取四张图片，进行随机抖动，然后将四张拼成一张图
代码如下：
import os
from random import sample
import numpy as np
from PIL import Image, ImageDraw
from utils.random_data import get_random_data, get_random_data_with_Mosaic
from utils.utils import convert_annotation, get_classes
#-----------------------------------------------------------------------------------#
# Origin_VOCdevkit_path 原始标签所在的路径
# Out_VOCdevkit_path 输出标签所在的路径
# 处理后的标签为灰度图，如果设置的值太小会看不见具体情况。
#-----------------------------------------------------------------------------------#
Origin_VOCdevkit_path = "VOCdevkit_Origin"
Out_VOCdevkit_path = "VOCdevkit"
#-----------------------------------------------------------------------------------#
# Out_Num 利用mixup生成多少组图片
# input_shape 生成的图片大小
#-----------------------------------------------------------------------------------#
Out_Num = 50
input_shape = [640, 640]
#-----------------------------------------------------------------------------------#
# 下面定义了xml里面的组成模块，无需改动。
#-----------------------------------------------------------------------------------#
headstr = """\
VOC
%s
My Database
COCO
flickr
NULL
NULL
company
%d
%d
%d
0
"""
objstr = """\
"""
tailstr = '''\
'''
if __name__ == "__main__":
Origin_JPEGImages_path = os.path.join(Origin_VOCdevkit_path, "VOC2007/JPEGImages")
Origin_Annotations_path = os.path.join(Origin_VOCdevkit_path, "VOC2007/Annotations")
Out_JPEGImages_path = os.path.join(Out_VOCdevkit_path, "VOC2007/JPEGImages")
Out_Annotations_path = os.path.join(Out_VOCdevkit_path, "VOC2007/Annotations")
if not os.path.exists(Out_JPEGImages_path):
os.makedirs(Out_JPEGImages_path)
if not os.path.exists(Out_Annotations_path):
os.makedirs(Out_Annotations_path)
#---------------------------#
# 遍历标签并赋值
#---------------------------#
xml_names = os.listdir(Origin_Annotations_path)
def write_xml(anno_path, jpg_pth, head, input_shape, boxes, unique_labels, tail):
f = open(anno_path, "w")
f.write(head%(jpg_pth, input_shape[0], input_shape[1], 3))
for i, box in enumerate(boxes):
f.write(objstr%(str(unique_labels[int(box[4])]), box[0], box[1], box[2], box[3]))
f.write(tail)
#########以上部分无需改动###########################################################################################
#------------------------------#
# 循环生成xml和jpg
#------------------------------#
for index in range(Out_Num):
#------------------------------#
# 获取4个图像与标签
#------------------------------#
sample_xmls = sample(xml_names, 4)
unique_labels = get_classes(sample_xmls, Origin_Annotations_path)
annotation_line = []
for xml in sample_xmls:
line = convert_annotation(os.path.join(Origin_JPEGImages_path, os.path.splitext(xml)[0] + '.jpeg'), os.path.join(Origin_Annotations_path, xml), unique_labels)
annotation_line.append(line)
#------------------------------#
# 合并mosaic
#------------------------------#
image_data, box_data = get_random_data_with_Mosaic(annotation_line, input_shape)
img = Image.fromarray(image_data.astype(np.uint8))
img.save(os.path.join(Out_JPEGImages_path, str(index) + '.jpeg'))
write_xml(os.path.join(Out_Annotations_path, str(index) + '.xml'), os.path.join(Out_JPEGImages_path, str(index) + '.jpeg'), \
headstr, input_shape, box_data, unique_labels, tailstr)
5.亮度，对比度调整
修改图像的一系列HSV参数，改变图像的亮度，色相，饱和度等数据
原图：
进行参数调整后：
import cv2
import math
import numpy as np
import os
import glob
import json
import shutil
import xml.etree.ElementTree as ET
from xml.etree.ElementTree import ElementTree, Element
def getColorImg(alpha,beta,img_path,img_write_path):
img = cv2.imread(img_path)
colored_img = np.uint8(np.clip((alpha * img + beta), 0, 255))
cv2.imwrite(img_write_path,colored_img)
def getColorAnno(anno_path,anno_write_path):
tree = ET.parse(anno_path)
tree.write(anno_write_path) # 保存修改后的XML文件
def color(alpha,beta,img_dir,anno_dir,img_write_dir,anno_write_dir):
if not os.path.exists(img_write_dir):
os.makedirs(img_write_dir)
if not os.path.exists(anno_write_dir):
os.makedirs(anno_write_dir)
img_names=os.listdir(img_dir)
for img_name in img_names:
img_path=os.path.join(img_dir,img_name)
img_write_path=os.path.join(img_write_dir,img_name[:-4]+'color'+str(int(alpha*10))+'.jpg')
#
anno_path=os.path.join(anno_dir,img_name[:-4]+'.xml')
anno_write_path = os.path.join(anno_write_dir, img_name[:-4]+'color'+str(int(alpha*10))+'.xml')
#
getColorImg(alpha,beta,img_path,img_write_path)
getColorAnno(anno_path,anno_write_path)
alphas=[0.3,0.5,1.2,1.6]
beta=10
img_dir='VOCdevkit_Origin/VOC2007/JPEGImages'
anno_dir='VOCdevkit_Origin/VOC2007/Annotations'
img_write_dir='H:\object-detection-augmentation\VOCdevkit\VOC2007\JPEGImages'
anno_write_dir='H:\object-detection-augmentation\VOCdevkit\VOC2007\Annotations'
for alpha in alphas:
color(alpha,beta,img_dir,anno_dir,img_write_dir,anno_write_dir)6.随机抖动
随机抖动是在对图像进行HSV变换之后，加上对图像的随机resize，裁剪，平移，旋转
原图：
随机抖动之后：
import os
from random import sample
import numpy as np
from PIL import Image, ImageDraw
from utils.random_data import get_random_data, get_random_data_with_MixUp
from utils.utils import convert_annotation, get_classes
#-----------------------------------------------------------------------------------#
# Origin_VOCdevkit_path 原始标签所在的路径
# Out_VOCdevkit_path 输出标签所在的路径
# 处理后的标签为灰度图，如果设置的值太小会看不见具体情况。
#-----------------------------------------------------------------------------------#
Origin_VOCdevkit_path = "VOCdevkit_Origin"
Out_VOCdevkit_path = "VOCdevkit"
#-----------------------------------------------------------------------------------#
# Out_Num 生成多少组图片
# input_shape 生成的图片大小
#-----------------------------------------------------------------------------------#
Out_Num = 30
input_shape = [640, 640]
#-----------------------------------------------------------------------------------#
# 下面定义了xml里面的组成模块，无需改动。
#-----------------------------------------------------------------------------------#
headstr = """\
VOC
%s
My Database
COCO
flickr
NULL
NULL
company
%d
%d
%d
0
"""
objstr = """\
"""
tailstr = '''\
'''
if __name__ == "__main__":
Origin_JPEGImages_path = os.path.join(Origin_VOCdevkit_path, "VOC2007/JPEGImages")
Origin_Annotations_path = os.path.join(Origin_VOCdevkit_path, "VOC2007/Annotations")
Out_JPEGImages_path = os.path.join(Out_VOCdevkit_path, "VOC2007/JPEGImages")
Out_Annotations_path = os.path.join(Out_VOCdevkit_path, "VOC2007/Annotations")
if not os.path.exists(Out_JPEGImages_path):
os.makedirs(Out_JPEGImages_path)
if not os.path.exists(Out_Annotations_path):
os.makedirs(Out_Annotations_path)
#---------------------------#
# 遍历标签并赋值
#---------------------------#
xml_names = os.listdir(Origin_Annotations_path)
def write_xml(anno_path, jpg_pth, head, input_shape, boxes, unique_labels, tail):
f = open(anno_path, "w")
f.write(head%(jpg_pth, input_shape[0], input_shape[1], 3))
for i, box in enumerate(boxes):
f.write(objstr%(str(unique_labels[int(box[4])]), box[0], box[1], box[2], box[3]))
f.write(tail)
#########以上部分无需改动###########################################################################################
#------------------------------#
# 循环生成xml和jpg
#------------------------------#
for index in range(Out_Num):
#------------------------------#
# 获取一个图像与标签
#------------------------------#
sample_xmls = sample(xml_names, 1)
unique_labels = get_classes(sample_xmls, Origin_Annotations_path)
jpg_name = os.path.join(Origin_JPEGImages_path, os.path.splitext(sample_xmls[0])[0] + '.jpg')
xml_name = os.path.join(Origin_Annotations_path, sample_xmls[0])
line = convert_annotation(jpg_name, xml_name, unique_labels)
#------------------------------#
# 各自数据增强
#------------------------------#
image_data, box_data = get_random_data(line, input_shape)
img = Image.fromarray(image_data.astype(np.uint8))
img.save(os.path.join(Out_JPEGImages_path, str(index) + '.jpg'))
write_xml(os.path.join(Out_Annotations_path, str(index) + '.xml'), os.path.join(Out_JPEGImages_path, str(index) + '.jpg'), \
headstr, input_shape, box_data, unique_labels, tailstr)