图像扩充用于图像目标检测

				版权声明：本文为博主原创文章，未经博主允许不得转载。					https://blog.csdn.net/wei_guo_xd/article/details/74199729				</div>
							            <link rel="stylesheet" href="https://csdnimg.cn/release/phoenix/template/css/ck_htmledit_views-d7e2a68c7c.css">
					<div class="htmledit_views" id="content_views">

常用的图像扩充方式有：

水平翻转，裁剪，视角变换，jpeg压缩，尺度变换，颜色变换，旋转

当用于分类数据集时，这些变换方法可以全部被使用，然而考虑到目标检测标注框的变换，我们选择如下几种方式用于目标检测数据集扩充：

jpeg压缩，尺度变换，颜色变换

这里，我们介绍一个图象变换包

http://lear.inrialpes.fr/people/paulin/projects/ITP/

这是项目主页，里面介绍了用于图像变换的基本方法，以及如何组合它们可以得到最好的效果，项目主页里同时带python程序。

里面的图像变换程序如下（用于windows下，用于目标检测时，做了一些修改）：

  
  
import os, sys, pdb, numpy
from PIL import Image,ImageChops,ImageOps,ImageDraw

#parameters used for the CVPR paper
NCROPS = 10
NHOMO = 8
JPG=[ 70, 50, 30]
ROTS = [ 3, 6, 9, 12, 15]
SCALES=[ 1.5** 0.5, 1.5, 1.5** 1.5, 1.5** 2, 1.5** 2.5]
#parameters computed on ILSVRC10 dataset
lcolor = [ 381688.61379382 , 4881.28307136, 2316.10313483]
pcolor = [[ -0.57848371, -0.7915924, 0.19681989],
[ -0.5795621 , 0.22908373, -0.78206676],
[ -0.57398987 , 0.56648223 , 0.59129816]]

#pre-generated gaussian values
alphas = [[ 0.004894 , 0.153527, -0.012182],
[ -0.058978, 0.114067, -0.061488],
[ 0.002428, -0.003576, -0.125031]]

def gen_colorimetry(i):
p1r = pcolor[ 0][ 0]
p1g = pcolor[ 1][ 0]
p1b = pcolor[ 2][ 0]
p2r = pcolor[ 0][ 1]
p2g = pcolor[ 1][ 1]
p2b = pcolor[ 2][ 1]
p3r = pcolor[ 0][ 2]
p3g = pcolor[ 1][ 2]
p3b = pcolor[ 2][ 2]

l1 = numpy.sqrt(lcolor[ 0])
l2 = numpy.sqrt(lcolor[ 1])
l3 = numpy.sqrt(lcolor[ 2])

if i<= 3:
alpha = alphas[i]
else:
numpy.random.seed(i* 3)
alpha = numpy.random.randn( 3, 0, 0.01)
a1 = alpha[ 0]
a2 = alpha[ 1]
a3 = alpha[ 2]

return (a1*l1*p1r + a2*l2*p2r + a3*l3*p3r,
a1*l1*p1g + a2*l2*p2g + a3*l3*p3g,
a1*l1*p1b + a2*l2*p2b + a3*l3*p3b)

def gen_crop(i,w,h):
numpy.random.seed( 4*i)
x0 = numpy.random.random()*(w/ 4)
y0 = numpy.random.random()*(h/ 4)
x1 = w - numpy.random.random()*(w/ 4)
y1 = h - numpy.random.random()*(h/ 4)

return (int(x0),int(y0),int(x1),int(y1))

def gen_homo(i,w,h):
if i== 0:
return ( 0, 0,int( 0.125*w),h,int( 0.875*w),h,w, 0)
elif i== 1:
return ( 0, 0,int( 0.25*w),h,int( 0.75*w),h,w, 0)
elif i== 2:
return ( 0,int( 0.125*h), 0,int( 0.875*h),w,h,w, 0)
elif i== 3:
return ( 0,int( 0.25*h), 0,int( 0.75*h),w,h,w, 0)
elif i== 4:
return (int( 0.125*w), 0, 0,h,w,h,int( 0.875*w), 0)
elif i== 5:
return (int( 0.25*w), 0, 0,h,w,h,int( 0.75*w), 0)
elif i== 6:
return ( 0, 0, 0,h,w,int( 0.875*h),w,int( 0.125*h))
elif i== 7:
return ( 0, 0, 0,h,w,int( 0.75*h),w,int( 0.25*h))
else:
assert False


def rot(image,angle,fname):
white = Image.new( 'L',image.size, "white")
wr = white.rotate(angle,Image.NEAREST,expand= 0)
im = image.rotate(angle,Image.BILINEAR,expand= 0)
try:
image.paste(im,wr)
except ValueError:
print >>sys.stderr, 'error: image do not match '+fname
return image

def gen_corner(n, w, h):
x0 = 0
x1 = w
y0 = 0
y1 = h

rat = 256 - 227

if n == 0: #center
x0 = (rat*w)/( 2* 256.0)
y0 = (rat*h)/( 2* 256.0)
x1 = w - (rat*w)/( 2* 256.0)
y1 = h - (rat*h)/( 2* 256.0)
elif n == 1:
x0 = (rat*w)/ 256.0
y0 = (rat*h)/ 256.0
elif n == 2:
x1 = w - (rat*w)/ 256.0
y0 = (rat*h)/ 256.0
elif n == 3:
x1 = w - (rat*w)/ 256.0
y1 = h - (rat*h)/ 256.0
else:
assert n== 4
x0 = (rat*w)/ 256.0
y1 = h - (rat*h)/ 256.0

return (int(x0),int(y0),int(x1),int(y1))

#the main fonction to call
#takes a image input path, a transformation and an output path and does the transformation
def gen_trans(imgfile,trans,outfile):
for trans in trans.split( '*'):
image = Image.open(imgfile)
w,h = image.size
if trans== "plain":
image.save(outfile, "JPEG",quality= 100)
elif trans== "flip":
ImageOps.mirror(image).save(outfile, "JPEG",quality= 100)
elif trans.startswith( "crop"):
c = int(trans[ 4:])
image.crop(gen_crop(c,w,h)).save(outfile, "JPEG",quality= 100)
elif trans.startswith( "homo"):
c = int(trans[ 4:])
image.transform((w,h),Image.QUAD,
gen_homo(c,w,h),
Image.BILINEAR).save(outfile, "JPEG",quality= 100)
elif trans.startswith( "jpg"):
image.save(outfile,quality=int(trans[ 3:]))
elif trans.startswith( "scale"):
scale = SCALES[int(trans.replace( "scale", ""))]
image.resize((int(w/scale),int(h/scale)),Image.BILINEAR).save(outfile, "JPEG",quality= 100)
elif trans.startswith( 'color'):
(dr,dg,db) = gen_colorimetry(int(trans[ 5]))
table = numpy.tile(numpy.arange( 256),( 3))
table[ : 256]+= (int)(dr)
table[ 256: 512]+= (int)(dg)
table[ 512: ]+= (int)(db)
image.convert( "RGB").point(table).save(outfile, "JPEG",quality= 100)
elif trans.startswith( 'rot-'):
angle =int(trans[ 4:])
for i in range(angle):
image = rot(image, -1,outfile)
image.save(outfile, "JPEG",quality= 100)
elif trans.startswith( 'rot'):
angle =int(trans[ 3:])
for i in range(angle):
image = rot(image, 1,outfile)
image.save(outfile, "JPEG",quality= 100)
elif trans.startswith( 'corner'):
i = int(trans[ 6:])
image.crop(gen_corner(i,w,h)).save(outfile, "JPEG",quality= 100)
else:
assert False, "Unrecognized transformation: "+trans
imgfile = outfile # in case we iterate


#Our 41 transformations used in the CVPR paper
def get_all_trans():
# transformations = (["plain","flip"]
# # +["crop%d"%i for i in range(NCROPS)]
# # +["homo%d"%i for i in range(NHOMO)]
# +["jpg%d"%i for i in JPG]
# +["scale0","scale1","scale2","scale3","scale4"]
# +["color%d"%i for i in range(3)]
# # +["rot-%d"%i for i in ROTS]
# # +["rot%d"%i for i in ROTS]
# )+["scale0","scale1","scale2","scale3","scale4"]
transformations=([ "plain"]
+ [ "jpg%d" % i for i in JPG]
+ [ "scale0", "scale1", "scale2", "scale3", "scale4"]
+ [ "color%d" % i for i in range( 3)])
return transformations

#transformations used at test time in deep architectures
def get_deep_trans():
return [ 'corner0', 'corner1', 'corner2', 'corner3', 'corner4', 'corner0*flip', 'corner1*flip', 'corner2*flip', 'corner3*flip', 'corner4*flip']

if __name__== "__main__":
inputpath = sys.argv[ 1]
name = [name for name in os.listdir(inputpath) if os.path.isfile(os.path.join(inputpath,name))]
#img_input = sys.argv[1]
outpath = sys.argv[ 2]
if len(sys.argv)>= 4:
trans = sys.argv[ 3]
if not trans.startswith( "["):
trans = [trans]
else:
trans = eval(trans)
else:
trans = get_all_trans()
print "Generating transformations and storing in %s"%(outpath)
for k in name:
for t in trans:
img_input=inputpath+ '\\'+k
gen_trans(img_input,t,outpath+ '\\%s_%s.jpg'%( ".".join(img_input.split( "\\")[ -1].split( ".")[: -1]),t))
#gen_trans(k, t, outpath + '\\%s_%s.jpg' % (".".join(k.split(".")[:-1]), t))
print "Finished. Transformations generated: %s"%( " ".join(trans))
  
  

这是变换前的图片：1._7_17.jpg

变换后的图片如下：

1_7_17_color0.jpg

1_7_17_color1.jpg

1_7_17_color2.jpg

1_7_17_jpg30.jpg

1_7_17_jog50.jpg

1_7_17_jpg70.jpg

1_7_17_scale0.jpg

1_7_17_scale1.jpg

1_7_17_scale2.jpg

1_7_17_scale3.jpg

1_7_17_scale4.jpg

用于目标检测时：XML标注文件也需要做相应修改，主要是针对尺度变换：

修改前的xml文件如下(1_7_17.jpg)：

   
   
<annotation>
<folder>spl </folder>
<filename>1_7_17.jpg </filename>
<source>
<database>The spl Database </database>
<annotation>The spl Database </annotation>
<image>spl </image>
<flickrid>0 </flickrid>
</source>
<owner>
<flickrid>spl </flickrid>
<name>xiaovv </name>
</owner>
<size>
<width>800 </width>
<height>800 </height>
<depth>3 </depth>
</size>
<segmented>0 </segmented>
<object>
<name>aeroplane </name>
<pose>Unspecified </pose>
<truncated>0 </truncated>
<difficult>0 </difficult>
<bndbox>
<xmin>151 </xmin>
<ymin>357 </ymin>
<xmax>212 </xmax>
<ymax>399 </ymax>
</bndbox>
</object>
<object>
<name>aeroplane </name>
<pose>Unspecified </pose>
<truncated>0 </truncated>
<difficult>0 </difficult>
<bndbox>
<xmin>134 </xmin>
<ymin>593 </ymin>
<xmax>193 </xmax>
<ymax>654 </ymax>
</bndbox>
</object>
</annotation>
   
   

修改后的xml文件如下(1_7_17_scale4.jpg)：

   
   
<?xml version="1.0" encoding="utf-8"?> <annotation>
<folder>spl </folder>
<filename>1_7_17_scale4.jpg </filename>
<source>
<database>The spl Database </database>
<annotation>The spl Database </annotation>
<image>spl </image>
<flickrid>0 </flickrid>
</source>
<owner>
<flickrid>spl </flickrid>
<name>xiaovv </name>
</owner>
<size>
<width>290 </width>
<height>290 </height>
<depth>3 </depth>
</size>
<segmented>0 </segmented>
<object>
<name>aeroplane </name>
<pose>Unspecified </pose>
<truncated>0 </truncated>
<difficult>0 </difficult>
<bndbox>
<xmin>54 </xmin>
<ymin>129 </ymin>
<xmax>76 </xmax>
<ymax>144 </ymax>
</bndbox>
</object>
<object>
<name>aeroplane </name>
<pose>Unspecified </pose>
<truncated>0 </truncated>
<difficult>0 </difficult>
<bndbox>
<xmin>48 </xmin>
<ymin>215 </ymin>
<xmax>70 </xmax>
<ymax>237 </ymax>
</bndbox>
</object>
</annotation>
   
   

修改xml文件的程序如下;

   
   
# -*- coding=utf-8 -*-
import os
import sys
import shutil
from xml.dom.minidom import Document
from xml.etree.ElementTree import ElementTree,Element
import xml.dom.minidom
JPG=[ 70, 50, 30]
SCALES=[ 1.5** 0.5, 1.5, 1.5** 1.5, 1.5** 2, 1.5** 2.5]

#产生变换后的xml文件
def gen_xml(xml_input,trans,outfile):
for trans in trans.split( '*'):
if trans== "plain" or trans.startswith( "jpg") or trans.startswith( 'color'): #如果是这几种变换，直接修改xml文件名就好
dom = xml.dom.minidom.parse(xml_input)
root = dom.documentElement
filenamelist = root.getElementsByTagName( 'filename')
filename = filenamelist[ 0]
c = str(filename.firstChild.data)
d = ".".join(outfile.split( "\\")[ -1].split( ".")[: -1]) + '.jpg'
filename.firstChild.data = d
f = open(outfile, 'w')
dom.writexml(f, encoding= 'utf-8')
elif trans.startswith( "scale"): #对于尺度变换，xml文件信息也需要改变
scale = SCALES[int(trans.replace( "scale", ""))]
dom=xml.dom.minidom.parse(xml_input)
root=dom.documentElement
filenamelist=root.getElementsByTagName( 'filename')
filename=filenamelist[ 0]
c=str(filename.firstChild.data)
d= ".".join(outfile.split( "\\")[ -1].split( ".")[: -1])+ '.jpg'
filename.firstChild.data=d
heightlist = root.getElementsByTagName( 'height')
height = heightlist[ 0]
a = int(height.firstChild.data)
b = str(int(a / scale))
height.firstChild.data = b
widthlist=root.getElementsByTagName( 'width')
width=widthlist[ 0]
a = int(width.firstChild.data)
b = str(int(a / scale))
width.firstChild.data=b
objectlist=root.getElementsByTagName( 'xmin')
for object in objectlist:
a=int(object.firstChild.data)
b=str(int(a/scale))
object.firstChild.data=b
objectlist = root.getElementsByTagName( 'ymin')
for object in objectlist:
a = int(object.firstChild.data)
b = str(int(a / scale))
object.firstChild.data = b
objectlist = root.getElementsByTagName( 'xmax')
for object in objectlist:
a = int(object.firstChild.data)
b = str(int(a / scale))
object.firstChild.data = b
objectlist = root.getElementsByTagName( 'ymax')
for object in objectlist:
a = int(object.firstChild.data)
b = str(int(a / scale))
object.firstChild.data = b
f=open(outfile, 'w')
dom.writexml(f,encoding= 'utf-8')
else:
assert False, "Unrecognized transformation: "+trans

#产生各种变换名
def get_all_trans():
transformations=([ "plain"]
+ [ "jpg%d" % i for i in JPG]
+ [ "scale0", "scale1", "scale2", "scale3", "scale4"]
+ [ "color%d" % i for i in range( 3)])
return transformations

if __name__== "__main__":
inputpath = sys.argv[ 1]
name = [name for name in os.listdir(inputpath) if os.path.isfile(os.path.join(inputpath,name))]
outpath = sys.argv[ 2]
if len(sys.argv)>= 4:
trans = sys.argv[ 3]
if not trans.startswith( "["):
trans = [trans]
else:
trans = eval(trans)
else:
trans = get_all_trans()
print "Generating transformations and storing in %s"%(outpath)
for k in name:
for t in trans:
xml_input=inputpath+ '\\'+k
gen_xml(xml_input,t,outpath+ '\\%s_%s.xml'%( ".".join(xml_input.split( "\\")[ -1].split( ".")[: -1]),t))
   
   

				版权声明：本文为博主原创文章，未经博主允许不得转载。					https://blog.csdn.net/wei_guo_xd/article/details/74199729				</div>
							            <link rel="stylesheet" href="https://csdnimg.cn/release/phoenix/template/css/ck_htmledit_views-d7e2a68c7c.css">
					<div class="htmledit_views" id="content_views">