图像扩充用于图像目标检测

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