1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
|
#!/usr/bin/env python
'''
SVM and KNearest digit recognition.
Sample loads a dataset of handwritten digits from '../data/digits.png'.
Then it trains a SVM and KNearest classifiers on it and evaluates
their accuracy.
Following preprocessing is applied to the dataset:
- Moment-based image deskew (see deskew())
- Digit images are split into 4 10x10 cells and 16-bin
histogram of oriented gradients is computed for each
cell
- Transform histograms to space with Hellinger metric (see [1] (RootSIFT))
[1] R. Arandjelovic, A. Zisserman
"Three things everyone should know to improve object retrieval"
http://www.robots.ox.ac.uk/~vgg/publications/2012/Arandjelovic12/arandjelovic12.pdf
'''
# Python 2/3 compatibility
from __future__ import print_function
# built-in modules
from multiprocessing.pool import ThreadPool
import cv2
import numpy as np
from numpy.linalg import norm
SZ = 20 # size of each digit is SZ x SZ
CLASS_N = 10
DIGITS_FN = 'samples/data/digits.png'
def split2d(img, cell_size, flatten=True):
h, w = img.shape[:2]
sx, sy = cell_size
cells = [np.hsplit(row, w//sx) for row in np.vsplit(img, h//sy)]
cells = np.array(cells)
if flatten:
cells = cells.reshape(-1, sy, sx)
return cells
def deskew(img):
m = cv2.moments(img)
if abs(m['mu02']) < 1e-2:
return img.copy()
skew = m['mu11']/m['mu02']
M = np.float32([[1, skew, -0.5*SZ*skew], [0, 1, 0]])
img = cv2.warpAffine(img, M, (SZ, SZ), flags=cv2.WARP_INVERSE_MAP | cv2.INTER_LINEAR)
return img
class StatModel(object):
def load(self, fn):
self.model.load(fn) # Known bug: https://github.com/opencv/opencv/issues/4969
def save(self, fn):
self.model.save(fn)
class KNearest(StatModel):
def __init__(self, k = 3):
self.k = k
self.model = cv2.ml.KNearest_create()
def train(self, samples, responses):
self.model.train(samples, cv2.ml.ROW_SAMPLE, responses)
def predict(self, samples):
retval, results, neigh_resp, dists = self.model.findNearest(samples, self.k)
return results.ravel()
class SVM(StatModel):
def __init__(self, C = 1, gamma = 0.5):
self.model = cv2.ml.SVM_create()
self.model.setGamma(gamma)
self.model.setC(C)
self.model.setKernel(cv2.ml.SVM_RBF)
self.model.setType(cv2.ml.SVM_C_SVC)
def train(self, samples, responses):
self.model.train(samples, cv2.ml.ROW_SAMPLE, responses)
def predict(self, samples):
return self.model.predict(samples)[1].ravel()
def evaluate_model(model, digits, samples, labels):
resp = model.predict(samples)
err = (labels != resp).mean()
confusion = np.zeros((10, 10), np.int32)
for i, j in zip(labels, resp):
confusion[int(i), int(j)] += 1
return err, confusion
def preprocess_simple(digits):
return np.float32(digits).reshape(-1, SZ*SZ) / 255.0
def preprocess_hog(digits):
samples = []
for img in digits:
gx = cv2.Sobel(img, cv2.CV_32F, 1, 0)
gy = cv2.Sobel(img, cv2.CV_32F, 0, 1)
mag, ang = cv2.cartToPolar(gx, gy)
bin_n = 16
bin = np.int32(bin_n*ang/(2*np.pi))
bin_cells = bin[:10,:10], bin[10:,:10], bin[:10,10:], bin[10:,10:]
mag_cells = mag[:10,:10], mag[10:,:10], mag[:10,10:], mag[10:,10:]
hists = [np.bincount(b.ravel(), m.ravel(), bin_n) for b, m in zip(bin_cells, mag_cells)]
hist = np.hstack(hists)
# transform to Hellinger kernel
eps = 1e-7
hist /= hist.sum() + eps
hist = np.sqrt(hist)
hist /= norm(hist) + eps
samples.append(hist)
return np.float32(samples)
from tests_common import NewOpenCVTests
class digits_test(NewOpenCVTests):
def load_digits(self, fn):
digits_img = self.get_sample(fn, 0)
digits = split2d(digits_img, (SZ, SZ))
labels = np.repeat(np.arange(CLASS_N), len(digits)/CLASS_N)
return digits, labels
def test_digits(self):
digits, labels = self.load_digits(DIGITS_FN)
# shuffle digits
rand = np.random.RandomState(321)
shuffle = rand.permutation(len(digits))
digits, labels = digits[shuffle], labels[shuffle]
digits2 = list(map(deskew, digits))
samples = preprocess_hog(digits2)
train_n = int(0.9*len(samples))
digits_train, digits_test = np.split(digits2, [train_n])
samples_train, samples_test = np.split(samples, [train_n])
labels_train, labels_test = np.split(labels, [train_n])
errors = list()
confusionMatrixes = list()
model = KNearest(k=4)
model.train(samples_train, labels_train)
error, confusion = evaluate_model(model, digits_test, samples_test, labels_test)
errors.append(error)
confusionMatrixes.append(confusion)
model = SVM(C=2.67, gamma=5.383)
model.train(samples_train, labels_train)
error, confusion = evaluate_model(model, digits_test, samples_test, labels_test)
errors.append(error)
confusionMatrixes.append(confusion)
eps = 0.001
normEps = len(samples_test) * 0.02
confusionKNN = [[45, 0, 0, 0, 0, 0, 0, 0, 0, 0],
[ 0, 57, 0, 0, 0, 0, 0, 0, 0, 0],
[ 0, 0, 59, 1, 0, 0, 0, 0, 1, 0],
[ 0, 0, 0, 43, 0, 0, 0, 1, 0, 0],
[ 0, 0, 0, 0, 38, 0, 2, 0, 0, 0],
[ 0, 0, 0, 2, 0, 48, 0, 0, 1, 0],
[ 0, 1, 0, 0, 0, 0, 51, 0, 0, 0],
[ 0, 0, 1, 0, 0, 0, 0, 54, 0, 0],
[ 0, 0, 0, 0, 0, 1, 0, 0, 46, 0],
[ 1, 1, 0, 1, 1, 0, 0, 0, 2, 42]]
confusionSVM = [[45, 0, 0, 0, 0, 0, 0, 0, 0, 0],
[ 0, 57, 0, 0, 0, 0, 0, 0, 0, 0],
[ 0, 0, 59, 2, 0, 0, 0, 0, 0, 0],
[ 0, 0, 0, 43, 0, 0, 0, 1, 0, 0],
[ 0, 0, 0, 0, 40, 0, 0, 0, 0, 0],
[ 0, 0, 0, 1, 0, 50, 0, 0, 0, 0],
[ 0, 0, 0, 0, 1, 0, 51, 0, 0, 0],
[ 0, 0, 1, 0, 0, 0, 0, 54, 0, 0],
[ 0, 0, 0, 0, 0, 0, 0, 0, 47, 0],
[ 0, 1, 0, 1, 0, 0, 0, 0, 1, 45]]
self.assertLess(cv2.norm(confusionMatrixes[0] - confusionKNN, cv2.NORM_L1), normEps)
self.assertLess(cv2.norm(confusionMatrixes[1] - confusionSVM, cv2.NORM_L1), normEps)
self.assertLess(errors[0] - 0.034, eps)
self.assertLess(errors[1] - 0.018, eps)
|
