目标检测图像增强

https://blog.csdn.net/wei_guo_xd/article/details/74199729

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）：

 

1. import os, sys, pdb, numpy

2. from PIL import Image,ImageChops,ImageOps,ImageDraw

3.  

4. #parameters used for the CVPR paper

5. NCROPS = 10

6. NHOMO = 8

7. JPG=[70,50,30]

8. ROTS = [3,6,9,12,15]

9. SCALES=[1.5**0.5,1.5,1.5**1.5,1.5**2,1.5**2.5]

10. #parameters computed on ILSVRC10 dataset

11. lcolor = [ 381688.61379382 , 4881.28307136, 2316.10313483]

12. pcolor = [[-0.57848371, -0.7915924, 0.19681989],

13. [-0.5795621 , 0.22908373, -0.78206676],

14. [-0.57398987 , 0.56648223 , 0.59129816]]

15.  

16. #pre-generated gaussian values

17. alphas = [[0.004894 , 0.153527, -0.012182],

18. [-0.058978, 0.114067, -0.061488],

19. [0.002428, -0.003576, -0.125031]]

20.  

21. def gen_colorimetry(i):

22. p1r = pcolor[0][0]

23. p1g = pcolor[1][0]

24. p1b = pcolor[2][0]

25. p2r = pcolor[0][1]

26. p2g = pcolor[1][1]

27. p2b = pcolor[2][1]

28. p3r = pcolor[0][2]

29. p3g = pcolor[1][2]

30. p3b = pcolor[2][2]

31.  

32. l1 = numpy.sqrt(lcolor[0])

33. l2 = numpy.sqrt(lcolor[1])

34. l3 = numpy.sqrt(lcolor[2])

35.  

36. if i<=3:

37. alpha = alphas[i]

38. else:

39. numpy.random.seed(i*3)

40. alpha = numpy.random.randn(3,0,0.01)

41. a1 = alpha[0]

42. a2 = alpha[1]

43. a3 = alpha[2]

44.  

45. return (a1*l1*p1r + a2*l2*p2r + a3*l3*p3r,

46. a1*l1*p1g + a2*l2*p2g + a3*l3*p3g,

47. a1*l1*p1b + a2*l2*p2b + a3*l3*p3b)

48.  

49. def gen_crop(i,w,h):

50. numpy.random.seed(4*i)

51. x0 = numpy.random.random()*(w/4)

52. y0 = numpy.random.random()*(h/4)

53. x1 = w - numpy.random.random()*(w/4)

54. y1 = h - numpy.random.random()*(h/4)

55.  

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

57.  

58. def gen_homo(i,w,h):

59. if i==0:

60. return (0,0,int(0.125*w),h,int(0.875*w),h,w,0)

61. elif i==1:

62. return (0,0,int(0.25*w),h,int(0.75*w),h,w,0)

63. elif i==2:

64. return (0,int(0.125*h),0,int(0.875*h),w,h,w,0)

65. elif i==3:

66. return (0,int(0.25*h),0,int(0.75*h),w,h,w,0)

67. elif i==4:

68. return (int(0.125*w),0,0,h,w,h,int(0.875*w),0)

69. elif i==5:

70. return (int(0.25*w),0,0,h,w,h,int(0.75*w),0)

71. elif i==6:

72. return (0,0,0,h,w,int(0.875*h),w,int(0.125*h))

73. elif i==7:

74. return (0,0,0,h,w,int(0.75*h),w,int(0.25*h))

75. else:

76. assert False

77.  
78.  

79. def rot(image,angle,fname):

80. white = Image.new('L',image.size,"white")

81. wr = white.rotate(angle,Image.NEAREST,expand=0)

82. im = image.rotate(angle,Image.BILINEAR,expand=0)

83. try:

84. image.paste(im,wr)

85. except ValueError:

86. print >>sys.stderr, 'error: image do not match '+fname

87. return image

88.  

89. def gen_corner(n, w, h):

90. x0 = 0

91. x1 = w

92. y0 = 0

93. y1 = h

94.  

95. rat = 256 - 227

96.  

97. if n == 0: #center

98. x0 = (rat*w)/(2*256.0)

99. y0 = (rat*h)/(2*256.0)

100. x1 = w - (rat*w)/(2*256.0)

101. y1 = h - (rat*h)/(2*256.0)

102. elif n == 1:

103. x0 = (rat*w)/256.0

104. y0 = (rat*h)/256.0

105. elif n == 2:

106. x1 = w - (rat*w)/256.0

107. y0 = (rat*h)/256.0

108. elif n == 3:

109. x1 = w - (rat*w)/256.0

110. y1 = h - (rat*h)/256.0

111. else:

112. assert n==4

113. x0 = (rat*w)/256.0

114. y1 = h - (rat*h)/256.0

115.  

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

117.  

118. #the main fonction to call

119. #takes a image input path, a transformation and an output path and does the transformation

120. def gen_trans(imgfile,trans,outfile):

121. for trans in trans.split('*'):

122. image = Image.open(imgfile)

123. w,h = image.size

124. if trans=="plain":

125. image.save(outfile,"JPEG",quality=100)

126. elif trans=="flip":

127. ImageOps.mirror(image).save(outfile,"JPEG",quality=100)

128. elif trans.startswith("crop"):

129. c = int(trans[4:])

130. image.crop(gen_crop(c,w,h)).save(outfile,"JPEG",quality=100)

131. elif trans.startswith("homo"):

132. c = int(trans[4:])

133. image.transform((w,h),Image.QUAD,

134. gen_homo(c,w,h),

135. Image.BILINEAR).save(outfile,"JPEG",quality=100)

136. elif trans.startswith("jpg"):

137. image.save(outfile,quality=int(trans[3:]))

138. elif trans.startswith("scale"):

139. scale = SCALES[int(trans.replace("scale",""))]

140. image.resize((int(w/scale),int(h/scale)),Image.BILINEAR).save(outfile,"JPEG",quality=100)

141. elif trans.startswith('color'):

142. (dr,dg,db) = gen_colorimetry(int(trans[5]))

143. table = numpy.tile(numpy.arange(256),(3))

144. table[ :256]+= (int)(dr)

145. table[256:512]+= (int)(dg)

146. table[512: ]+= (int)(db)

147. image.convert("RGB").point(table).save(outfile,"JPEG",quality=100)

148. elif trans.startswith('rot-'):

149. angle =int(trans[4:])

150. for i in range(angle):

151. image = rot(image,-1,outfile)

152. image.save(outfile,"JPEG",quality=100)

153. elif trans.startswith('rot'):

154. angle =int(trans[3:])

155. for i in range(angle):

156. image = rot(image,1,outfile)

157. image.save(outfile,"JPEG",quality=100)

158. elif trans.startswith('corner'):

159. i = int(trans[6:])

160. image.crop(gen_corner(i,w,h)).save(outfile,"JPEG",quality=100)

161. else:

162. assert False, "Unrecognized transformation: "+trans

163. imgfile = outfile # in case we iterate

164.  
165.  

166. #Our 41 transformations used in the CVPR paper

167. def get_all_trans():

168. # transformations = (["plain","flip"]

169. # # +["crop%d"%i for i in range(NCROPS)]

170. # # +["homo%d"%i for i in range(NHOMO)]

171. # +["jpg%d"%i for i in JPG]

172. # +["scale0","scale1","scale2","scale3","scale4"]

173. # +["color%d"%i for i in range(3)]

174. # # +["rot-%d"%i for i in ROTS]

175. # # +["rot%d"%i for i in ROTS]

176. # )+["scale0","scale1","scale2","scale3","scale4"]

177. transformations=(["plain"]

178. + ["jpg%d" % i for i in JPG]

179. + ["scale0", "scale1", "scale2", "scale3", "scale4"]

180. + ["color%d" % i for i in range(3)])

181. return transformations

182.  

183. #transformations used at test time in deep architectures

184. def get_deep_trans():

185. return ['corner0','corner1','corner2','corner3','corner4','corner0*flip','corner1*flip','corner2*flip','corner3*flip','corner4*flip']

186.  

187. if __name__=="__main__":

188. inputpath = sys.argv[1]

189. name = [name for name in os.listdir(inputpath) if os.path.isfile(os.path.join(inputpath,name))]

190. #img_input = sys.argv[1]

191. outpath = sys.argv[2]

192. if len(sys.argv)>= 4:

193. trans = sys.argv[3]

194. if not trans.startswith("["):

195. trans = [trans]

196. else:

197. trans = eval(trans)

198. else:

199. trans = get_all_trans()

200. print "Generating transformations and storing in %s"%(outpath)

201. for k in name:

202. for t in trans:

203. img_input=inputpath+'\\'+k

204. gen_trans(img_input,t,outpath+'\\%s_%s.jpg'%(".".join(img_input.split("\\")[-1].split(".")[:-1]),t))

205. #gen_trans(k, t, outpath + '\\%s_%s.jpg' % (".".join(k.split(".")[:-1]), t))

206. print "Finished. Transformations generated: %s"%(" ".join(trans))

修改xml文件的程序如下;

 

1. # -*- coding=utf-8 -*-

2. import os

3. import sys

4. import shutil

5. from xml.dom.minidom import Document

6. from xml.etree.ElementTree import ElementTree,Element

7. import xml.dom.minidom

8. JPG=[70,50,30]

9. SCALES=[1.5**0.5,1.5,1.5**1.5,1.5**2,1.5**2.5]

10.  

11. #产生变换后的xml文件

12. def gen_xml(xml_input,trans,outfile):

13. for trans in trans.split('*'):

14. 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

15. dom = xml.dom.minidom.parse(xml_input)

16. root = dom.documentElement

17. filenamelist = root.getElementsByTagName('filename')

18. filename = filenamelist[0]

19. c = str(filename.firstChild.data)

20. d = ".".join(outfile.split("\\")[-1].split(".")[:-1]) + '.jpg'

21. filename.firstChild.data = d

22. f = open(outfile, 'w')

23. dom.writexml(f, encoding='utf-8')

24. elif trans.startswith("scale"):#对于尺度变换，xml文件信息也需要改变

25. scale = SCALES[int(trans.replace("scale", ""))]

26. dom=xml.dom.minidom.parse(xml_input)

27. root=dom.documentElement

28. filenamelist=root.getElementsByTagName('filename')

29. filename=filenamelist[0]

30. c=str(filename.firstChild.data)

31. d=".".join(outfile.split("\\")[-1].split(".")[:-1])+'.jpg'

32. filename.firstChild.data=d

33. heightlist = root.getElementsByTagName('height')

34. height = heightlist[0]

35. a = int(height.firstChild.data)

36. b = str(int(a / scale))

37. height.firstChild.data = b

38. widthlist=root.getElementsByTagName('width')

39. width=widthlist[0]

40. a = int(width.firstChild.data)

41. b = str(int(a / scale))

42. width.firstChild.data=b

43. objectlist=root.getElementsByTagName('xmin')

44. for object in objectlist:

45. a=int(object.firstChild.data)

46. b=str(int(a/scale))

47. object.firstChild.data=b

48. objectlist = root.getElementsByTagName('ymin')

49. for object in objectlist:

50. a = int(object.firstChild.data)

51. b = str(int(a / scale))

52. object.firstChild.data = b

53. objectlist = root.getElementsByTagName('xmax')

54. for object in objectlist:

55. a = int(object.firstChild.data)

56. b = str(int(a / scale))

57. object.firstChild.data = b

58. objectlist = root.getElementsByTagName('ymax')

59. for object in objectlist:

60. a = int(object.firstChild.data)

61. b = str(int(a / scale))

62. object.firstChild.data = b

63. f=open(outfile,'w')

64. dom.writexml(f,encoding='utf-8')

65. else:

66. assert False, "Unrecognized transformation: "+trans

67.  

68. #产生各种变换名

69. def get_all_trans():

70. transformations=(["plain"]

71. + ["jpg%d" % i for i in JPG]

72. + ["scale0", "scale1", "scale2", "scale3", "scale4"]

73. + ["color%d" % i for i in range(3)])

74. return transformations

75.  

76. if __name__=="__main__":

77. inputpath = sys.argv[1]

78. name = [name for name in os.listdir(inputpath) if os.path.isfile(os.path.join(inputpath,name))]

79. outpath = sys.argv[2]

80. if len(sys.argv)>= 4:

81. trans = sys.argv[3]

82. if not trans.startswith("["):

83. trans = [trans]

84. else:

85. trans = eval(trans)

86. else:

87. trans = get_all_trans()

88. print "Generating transformations and storing in %s"%(outpath)

89. for k in name:

90. for t in trans:

91. xml_input=inputpath+'\\'+k

92. gen_xml(xml_input,t,outpath+'\\%s_%s.xml'%(".".join(xml_input.split("\\")[-1].split(".")[:-1]),t))