简述人脸特异性识别&&一个基于LBP和SVM的人脸识别小例子

  1 ###################################################
  2 #
  3 #
  4 #    NO LONGER USEFUL....
  5 #
  6 #
  7 ###################################################
  8 def lbp_ave(img, radius, x, y):
  9     pix_count = 0
 10     accu = 0
 11     for i in range(x, x+radius):
 12         for j in range(y, y+radius):
 13             if(i >= 0 and i < img.shape[0] and j>=0 and j < img.shape[1]):
 14                 pix_count += 1
 15                 accu += img[i][j]
 16     if(pix_count != 0):
 17         return accu/pix_count
 18     return 0
 19 
 20 ###################################################
 21 #
 22 #
 23 #    Now dolbp is a simple eight-point lbp calculating function
 24 #
 25 #    it will be updated to multi-point scalable-radius version later
 26 #
 27 #
 28 ###################################################
 29 def dolbp(img, x, y, radius, numsamplepoints):
 30     lbpval = 0
 31     threshold = img[x][y]
 32     if threshold <= img[x - 1][y]:
 33         lbpval += 128
 34     if threshold <= img[x - 1][y - 1]:
 35         lbpval += 64
 36     if threshold <= img[x][y - 1]:
 37         lbpval += 32
 38     if threshold <= img[x + 1][y - 1]:
 39         lbpval += 16
 40     if threshold <= img[x + 1][y]:
 41         lbpval += 8
 42     if threshold <= img[x + 1][y + 1]:
 43         lbpval += 4
 44     if threshold <= img[x][y + 1]:
 45         lbpval += 2
 46     if threshold <= img[x - 1][y + 1]:
 47         lbpval += 1
 48 
 49     return lbpval
 50 def isuniform(val, numpoints):
 51     bit = 0
 52     lastbit = 0
 53     changecount = 0
 54     if val & (1 << (numpoints-1)):
 55         bit = 1
 56     else:
 57         bit = 0
 58     
 59     for i in range(0, numpoints-1)[::-1]:
 60         if val & (1 << i):
 61             bit = 1
 62         else:
 63             bit = 0
 64         if bit != lastbit:
 65             changecount += 1
 66         lastbit = bit
 67     
 68     if changecount <= 2:
 69         return True
 70     else:
 71         return False
 72 
 73 def lbpfe_uniform(histogram, numpoints):
 74     notuniformcount = 0
 75     histogram_uniformed = []
 76     for i in range(2**numpoints):
 77         if histogram[i] >= 0:#make sure the number of this bin is minimal
 78             if isuniform(i, numpoints) == True:
 79                 histogram_uniformed.append(histogram[i])
 80             else:
 81                 notuniformcount += histogram[i]
 82     return histogram_uniformed
 83 
 84 #if a bin
 85 def lbpfe_minimalize(histogram):
 86     return histogram
 87 
 88 def lbpfe_normalize(histogram):
 89     accu = 0
 90     normalized = []
 91     for i in histogram:
 92         accu += i
 93     if accu == 0:
 94         for i in histogram:
 95             normalized.append(float(0))
 96     else:
 97         for i in histogram:
 98             normalized.append(float(i)/accu)
 99     return normalized
100 
101 #x,y is the position of the cell‘s lower-left pixel
102 def lbpfe_calccell(img, x, y, cellsize, isminimalized, radius, numsamplepoints):
103     #Generate the histogram list
104     histogram_raw = []
105     #add zeros to the list
106     for i in range(2**numsamplepoints):
107         histogram_raw.append(0)
108     #do advanced LBP for each pixel in the cell
109     for i in range(x, x+cellsize-1):
110         for j in range(y, y+cellsize-1):
111             histogram_raw[dolbp(img, i, j, radius, numsamplepoints)] += 1
112             
113     #make it minimal
114     # if a bin iis meaningless, it will be set to -1
115     if isminimalized == True:
116         histogram_minimalized = lbpfe_minimalize(histogram_raw)
117     else:
118         histogram_minimalized = histogram_raw
119     #Make it uniform
120     histogram_uniformed = lbpfe_uniform(histogram_minimalized, numsamplepoints)
121     #Normalize it!
122     histogram_normalized = lbpfe_normalize(histogram_uniformed)
123     # Well done!!!!!!!!
124     return histogram_normalized
125 ###################################################
126 #
127 #    LBP feature extraction function.......
128 #
129 #    img: input greyscale image array
130 #    point: vectors of interest [[x1,y1],[x2,y2],......]
131 #    cellsize: ^-^
132 #    isminimalized: 
133 #
134 #    there are only 58 uniform modes, so there are 59 categories
135 #
136 #    the cells is numbered like this:
137 #
138 #    ^ 13   14   15   16
139 #    |  9   10   11   12
140 #    |  5    6    7    8
141 #    y  1    2    3    4
142 #       x------------->
143 #
144 ###################################################
145 def lbpfe(img, points, cellsize, isminimalized=False, radius=1.0, numsamplepoints=8):
146     histogram_combined = []
147     histogram_nomalized = []
148     #for each interest point:
149     for point in points:
150         #for each cell:
151         histogram_nomalized = []
152         for i in [-2, -1, 0, 1]:
153             for j in [-2, -1, 0, 1]:
154                 histogram_nomalized.extend(lbpfe_calccell(img, int(point[0]) + i * cellsize, int(point[1]) + j * cellsize, cellsize, isminimalized, radius, numsamplepoints))
155                 #print i, j, ‘len h_n=‘, len(histogram_nomalized)
156         #store the result of each point to a list
157         histogram_combined.extend(histogram_nomalized)
158     return histogram_combined