[数据集处理]数据增强2
在[数据集处理]数据增强1方法上的提升，灵感来自yolo4的马赛克图像拼接方法的数据增强，该数据增强方法带来了以下优点：
（1）变相增加了bach_size；
（2）增加了背景复杂度；
（3）获得了多尺度的目标；
。。。。。。
import cv2
import os
import sys
import re
import xml.etree.ElementTree as ET
from PIL import Image
imgreadpath = './img/'
imgwritepath = './f_img/'
xmlreadpath = './xml/'
xmlwritepath = './f_xml/'
allimgreadpath = './allimg/'
ximgwritepath = './ximg/'
allxmlreadpath = './allxml/'
xxmlwritepath = './xxml/'
def flitimg(imgname):
image = cv2.imread(imgname)
name = imgname.split('/')[-1].split('.')[-2]
image_f = cv2.flip(image, 1) #1：水平翻转
cv2.imwrite(imgwritepath + 'f' + name + '.jpg',image_f)
def flitxml(xmlname):
bwidth = ''
bheight = ''
bdepth = ''
text=open(xmlname).read()
f_test = open(xmlwritepath + 'f' + xmlname.split("/")[-1].split('.')[-2] + '.xml', 'w')
text=re.sub(u"[\x00-\x08\x0b-\x0c\x0e-\x1f]+",u"",text)
root=ET.fromstring(text)
if root.findtext('folder') == ' ':
print("no folder filename:",filename)
print('A')
else:
folder = root.findtext('folder')
if root.findtext('filename') == ' ':
print("no filename filename:",filename)
print('B')
else:
filename = root.findtext('filename')
# if root.findall('size') == []:
# print("no size filename:",filename)
# print('C')
# else:
# size = root.findall('size')
# print(size)
# size = size[0]
# bwidth = size.findtext('width')
# bheight = size.findtext('height')
# bdepth = size.findtext('depth')
img = Image.open(imgwritepath + 'f' + xmlname.split("/")[-1].split('.')[-2] + '.jpg')
bwidth, bheight = img.size
bdepth =3
f_test.write('\n')
f_test.write(' ' + folder + ' \n')
f_test.write(' ' + filename + ' \n')
f_test.write(' \n')
f_test.write(' ' + str(bwidth) + ' \n')
f_test.write(' ' + str(bheight) + ' \n')
f_test.write(' ' + str(bdepth) + ' \n')
f_test.write(' \n')
if root.findall('object') == []:
print("no object filename:",filename)
else:
for object in root.findall('object'):
label = object.findtext('name')
x1 = object.findtext('bndbox/xmin')
y1 = object.findtext('bndbox/ymin')
x2 = object.findtext('bndbox/xmax')
y2 = object.findtext('bndbox/ymax')
f_test.write(' \n')
f_test.write(' \n')
f_test.close()
def img4to1(image_names):
# image_names 是一个含有四个图片名的list
to_image = Image.new('RGB', (IMAGE_COLUMN * IMAGE_SIZE, IMAGE_ROW * IMAGE_SIZE))
for y in range(1, IMAGE_ROW + 1):
for x in range(1, IMAGE_COLUMN + 1):
from_image = Image.open(allimgreadpath + image_names[IMAGE_COLUMN * (y - 1) + x - 1]).resize((IMAGE_SIZE, IMAGE_SIZE),Image.ANTIALIAS)
to_image.paste(from_image, ((x - 1) * IMAGE_SIZE, (y - 1) * IMAGE_SIZE))
to_image.save(ximgwritepath + 'ximg' + str(NUM) + '.jpg')
return 'ximg' + str(NUM) + '.jpg'
def xml4to1(image_names,xxmlname):
f_test = open(xxmlwritepath + xxmlname.split('.')[-2] + '.xml', 'w')
f_test.write('\n')
f_test.write(' ' + ximgwritepath + ' \n')
f_test.write(' ' + xxmlname + ' \n')
f_test.write(' \n')
f_test.write(' ' + str(IMAGE_COLUMN*IMAGE_SIZE) + ' \n')
f_test.write(' ' + str(IMAGE_ROW*IMAGE_SIZE) + ' \n')
f_test.write(' ' + str(3) + ' \n')
f_test.write(' \n')
for i in range(len(image_names)):
if i == 0:
img = Image.open(allimgreadpath + image_names[i])
width, height = img.size
scalewidth = IMAGE_SIZE/width
scaleheight = IMAGE_SIZE/height
text=open(allxmlreadpath + image_names[i].split('.')[-2] + '.xml').read()
text=re.sub(u"[\x00-\x08\x0b-\x0c\x0e-\x1f]+",u"",text)
root=ET.fromstring(text)
if root.findall('object') == []:
print("no object filename:",filename)
else:
for object in root.findall('object'):
label = object.findtext('name')
x1 = object.findtext('bndbox/xmin')
y1 = object.findtext('bndbox/ymin')
x2 = object.findtext('bndbox/xmax')
y2 = object.findtext('bndbox/ymax')
newx1 = int(int(x1)*scalewidth)
newy1 = int(int(y1)*scaleheight)
newx2 = int(int(x2)*scalewidth)
newy2 = int(int(y2)*scaleheight)
# if (newx2-newx1)>MINISCALE and (newy2-newy1)>MINISCALE:
if (newy2-newy1)>MINISCALE:
f_test.write(' \n')
else:
continue
elif i == 1:
img = Image.open(allimgreadpath + image_names[i])
width, height = img.size
scalewidth = IMAGE_SIZE/width
scaleheight = IMAGE_SIZE/height
text=open(allxmlreadpath + image_names[i].split('.')[-2] + '.xml').read()
text=re.sub(u"[\x00-\x08\x0b-\x0c\x0e-\x1f]+",u"",text)
root=ET.fromstring(text)
if root.findall('object') == []:
print("no object filename:",filename)
else:
for object in root.findall('object'):
label = object.findtext('name')
x1 = object.findtext('bndbox/xmin')
y1 = object.findtext('bndbox/ymin')
x2 = object.findtext('bndbox/xmax')
y2 = object.findtext('bndbox/ymax')
newx1 = int(int(x1)*scalewidth)
newy1 = int(int(y1)*scaleheight)
newx2 = int(int(x2)*scalewidth)
newy2 = int(int(y2)*scaleheight)
# if (newx2-newx1)>MINISCALE and (newy2-newy1)>MINISCALE:
if (newy2-newy1)>MINISCALE:
f_test.write(' \n')
else:
continue
elif i == 2:
img = Image.open(allimgreadpath + image_names[i])
width, height = img.size
scalewidth = IMAGE_SIZE/width
scaleheight = IMAGE_SIZE/height
text=open(allxmlreadpath + image_names[i].split('.')[-2] + '.xml').read()
text=re.sub(u"[\x00-\x08\x0b-\x0c\x0e-\x1f]+",u"",text)
root=ET.fromstring(text)
if root.findall('object') == []:
print("no object filename:",filename)
else:
for object in root.findall('object'):
label = object.findtext('name')
x1 = object.findtext('bndbox/xmin')
y1 = object.findtext('bndbox/ymin')
x2 = object.findtext('bndbox/xmax')
y2 = object.findtext('bndbox/ymax')
newx1 = int(int(x1)*scalewidth)
newy1 = int(int(y1)*scaleheight)
newx2 = int(int(x2)*scalewidth)
newy2 = int(int(y2)*scaleheight)
# if (newx2-newx1)>MINISCALE and (newy2-newy1)>MINISCALE:
if (newy2-newy1)>MINISCALE:
f_test.write(' \n')
else:
continue
elif i == 3:
img = Image.open(allimgreadpath + image_names[i])
width, height = img.size
scalewidth = IMAGE_SIZE/width
scaleheight = IMAGE_SIZE/height
text=open(allxmlreadpath + image_names[i].split('.')[-2] + '.xml').read()
text=re.sub(u"[\x00-\x08\x0b-\x0c\x0e-\x1f]+",u"",text)
root=ET.fromstring(text)
if root.findall('object') == []:
print("no object filename:",filename)
else:
for object in root.findall('object'):
label = object.findtext('name')
x1 = object.findtext('bndbox/xmin')
y1 = object.findtext('bndbox/ymin')
x2 = object.findtext('bndbox/xmax')
y2 = object.findtext('bndbox/ymax')
newx1 = int(int(x1)*scalewidth)
newy1 = int(int(y1)*scaleheight)
newx2 = int(int(x2)*scalewidth)
newy2 = int(int(y2)*scaleheight)
# if (newx2-newx1)>MINISCALE and (newy2-newy1)>MINISCALE:
if (newy2-newy1)>MINISCALE:
f_test.write(' \n')
else:
continue
f_test.write(' \n')
f_test.close()
if __name__ == '__main__':
# # 图片镜像翻转
# imgnames = os.listdir(imgreadpath)
# for imgname in imgnames:
# imgname = imgreadpath + imgname
# flitimg(imgname)
# # 图片镜像翻转中对应的xml标注内容对应翻转
# xmlnames = os.listdir(xmlreadpath)
# for xmlname in xmlnames:
# xmlname = xmlreadpath + xmlname
# flitxml(xmlname)
# # 图片4合1 对应的xml标注内容同时4合1
allimgnames = os.listdir(allimgreadpath)
NUM = 0
IMAGE_ROW = 2
IMAGE_COLUMN = 2
IMAGE_SIZE = 320
MINISCALE = 10
for image_names in [allimgnames[i:i+4] for i in range(0,len(allimgnames),4)]:
NUM += 1
xxmlname = img4to1(image_names)
print(xxmlname)
xml4to1(image_names,xxmlname)
原创脚本，转载请注明出处
希望对你有所帮助
(•̀ᴗ•́)و ̑̑点赞