-
Notifications
You must be signed in to change notification settings - Fork 0
/
filters.py
191 lines (144 loc) · 6.74 KB
/
filters.py
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
import numpy as np
from scipy import ndimage
class ConvOp(object):
'''Base class for a convolutional operation'''
def __init__(self, name, shape=None):
assert(np.sum(np.mod(np.array(shape), 2) - 1) == 0)
self.name = name
self.shape = shape
def apply(self, target):
raise NotImplementedError('Not implemented in base class')
class ElementWiseMultiplyConvOp(ConvOp):
def __init__(self, weights, name='ElementWiseMultiplyConvOp'):
super(ElementWiseMultiplyConvOp, self).__init__(name, weights.shape)
self.w = weights
def apply(self, target):
return np.sum(self.w * target)
class UniformBlurConvOp(ElementWiseMultiplyConvOp):
def __init__(self, blur_size=3, name='UniformBlurConvOp'):
weight = np.ones((blur_size, blur_size))
super(UniformBlurConvOp, self).__init__(weight, name)
class BlobDetector(ElementWiseMultiplyConvOp):
def __init__(self, name='BlobDetector'):
weights_size = self.get_filter_size()
weights = self.get_dispersion(weights_size)
super(BlobDetector, self).__init__(weights, name)
@staticmethod
def get_x_coord_mat(shape):
return np.tile(np.abs(np.arange(shape[1]) - np.floor(shape[1] / 2)), (shape[1], 1))
@staticmethod
def get_y_coord_mat(shape):
return np.tile(np.abs(np.arange(shape[0]).reshape(shape[0], 1) - np.floor(shape[0] / 2)), (1, shape[0]))
@staticmethod
def get_coord_distance_mat(shape):
D1 = np.tile(
np.abs(np.arange(shape[1]) - np.floor(shape[1] / 2)), (shape[1], 1))
D2 = np.tile(np.abs(np.arange(shape[0]).reshape(
shape[0], 1) - np.floor(shape[0] / 2)), (1, shape[0]))
D = np.power(np.power(D1, 2.) + np.power(D2, 2.), 0.5)
return D
def get_dispersion(self, shape):
raise NotImplementedError('Not implemented in base class')
def get_filter_size(self):
raise NotImplementedError('Not implemented in base class')
class LaplacianBlobDetector(BlobDetector):
def __init__(self, spread, name='LaplacianBlobDetector'):
self.spread = spread
super(LaplacianBlobDetector, self).__init__(name)
def get_dispersion(self, shape):
D = self.get_coord_distance_mat(shape) ** 2.0
scale2 = self.spread ** 2.0
disp = np.power(np.pi * scale2, -1.)
disp = disp * ((D / (2. * scale2)) - 1.)
disp = disp * np.exp(-D / (2. * scale2))
return disp
def get_filter_size(self):
fsize = int(self.spread * 6)
if not (fsize & 0x1):
fsize += 1
return (fsize, fsize)
class ConvFilter(object):
'''Base convolutional filter class'''
def __init__(self, conv_op, name='ConvFilter', extender=PerpendicularEdgeExtender()):
self.name = name
self.c = conv_op
self.e = extender
def apply(self, im):
im_e = self.e.apply(im, self.c.shape)
res = np.zeros(im_e.shape)
hw = np.floor((self.c.shape[0]) / 2).astype(int)
hh = np.floor((self.c.shape[1]) / 2).astype(int)
for i in range(self.c.shape[0], self.c.shape[0] + im.shape[0]):
for j in range(self.c.shape[1], self.c.shape[1] + im.shape[1]):
res[i, j] = self.c.apply(
im_e[i - hw:i + hw + 1, j - hh:j + hh + 1])
return res[self.c.shape[0]:-self.c.shape[0], self.c.shape[1]:-self.c.shape[1]]
class IterativeConvFilter(ConvFilter):
def __init__(self, conv_ops, name='IterativeConvFilter', extender=PerpendicularEdgeExtender()):
super(IterativeConvFilter, self).__init__(None, name, extender)
self.cs = conv_ops
self.max_dims = (max(map(lambda c: c.shape[0], conv_ops)), max(
map(lambda c: c.shape[1], conv_ops)))
def apply(self, im):
im_e = self.e.apply(im, self.max_dims)
def apply_on_im_e(c):
return self.apply_conv_op(im_e, c)
return list(map(apply_on_im_e, self.cs))
def apply_conv_op(self, im, c):
res = np.zeros(im.shape)
hw = np.floor((c.shape[0]) / 2).astype(int)
hh = np.floor((c.shape[1]) / 2).astype(int)
for i in range(c.shape[0], im.shape[0] - c.shape[0]):
for j in range(c.shape[1], im.shape[1] - c.shape[1]):
res[i, j] = c.apply(im[i - hw:i + hw + 1, j - hh:j + hh + 1])
return res[self.max_dims[0]:-self.max_dims[0], self.max_dims[1]:-self.max_dims[1]]
class HarrisDetector(object):
def __init__(self, sigma_i, sigma_d, alpha):
self.sigma_i = sigma_i
self.sigma_d = sigma_d
self.alpha = alpha
@staticmethod
def get_grads_xy(X, mode='reflect', cval=0.):
'''Compute gradients using 0-padding'''
# N.B.: using ndimage for speed, was too slow with numpy gradient
Gx = ndimage.sobel(X, axis=1, mode=mode, cval=cval)
Gy = ndimage.sobel(X, axis=0, mode=mode, cval=cval)
return Gx, Gy
def get_M(self, X, mode='reflect', cval=0.):
'''Get the Harris-Laplace scale-adapted second moment matrix'''
# Compute the gaussian smoothed image
# N.B.: using ndimage for speed; could replace with gaussian filter
# class above
G = ndimage.gaussian_filter(X, self.sigma_d, mode=mode, cval=cval)
# Compute derivatives of gaussian-smoothed image in x and y directions
Lx, Ly = self.get_grads_xy(G, mode=mode, cval=cval)
# Compute second derivatives from first derivatives in x and y
# directions
Lxx, Lxy = self.get_grads_xy(Lx, mode=mode, cval=cval)
_, Lyy = self.get_grads_xy(Ly, mode=mode, cval=cval)
# Convolve each second derivative matrix with the integration gaussian
# N.B.: using ndimage for speed; could replace with gaussian filter
# class above
Lxx = ndimage.gaussian_filter(Lxx, self.sigma_i, mode=mode, cval=cval)
Lxy = ndimage.gaussian_filter(Lxy, self.sigma_i, mode=mode, cval=cval)
Lyy = ndimage.gaussian_filter(Lyy, self.sigma_i, mode=mode, cval=cval)
# Get an empty matrix and store the second derivatives
M = np.empty((*X.shape, 2, 2))
M[:, :, 0, 0] = Lxx
M[:, :, 1, 0] = Lxy
M[:, :, 0, 1] = Lxy
M[:, :, 1, 1] = Lyy
# Apply scale correction
M = (self.sigma_d ** 2.) * M
return M
def get_cornerness_from_M(self, X):
'''Compute the cornerness metric from a second moment matrix M'''
assert(X.ndim == 4)
det = (X[:, :, 0, 0] * X[:, :, 1, 1]) - (X[:, :, 0, 1] * X[:, :, 1, 0])
tr = X[:, :, 0, 0] + X[:, :, 1, 1]
cornerness = det - (self.alpha * (tr ** 2.))
return cornerness
def harris(self, X, mode='reflect', cval=0.):
M = self.get_M(X, mode=mode, cval=cval)
C = self.get_cornerness_from_M(M)
return C