-
Notifications
You must be signed in to change notification settings - Fork 0
/
extract_feature.py
162 lines (140 loc) · 7 KB
/
extract_feature.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
# -*- coding: utf-8 -*-
"""
Created on Tue Jul 21 09:35:37 2020
@author: CarrieLai
"""
import cv2
import numpy as np
import math
import os
############################ Hog descriptor #############################
class HoG(object):
def __init__(self, img, block_size, block_stride, cell_size, bin_num):
self.img = img
self.img_h = img.shape[1] # h, horizontal
self.img_v = img.shape[0] # v, vertical
self.block_size_h = block_size[1]
self.block_size_v = block_size[0]
self.block_stride_h = block_stride[0]
self.block_stride_v = block_stride[1]
self.cell_size_h = cell_size[1]
self.cell_size_v = cell_size[0]
self.bin_num = bin_num
self.bin_unit = 360 / bin_num
def gradient(self):
self.img = np.sqrt(self.img / np.max(self.img)) # gamma normalization
dx = cv2.Sobel(self.img, cv2.CV_64F, 1, 0, ksize=5)
dy = cv2.Sobel(self.img, cv2.CV_64F, 0, 1, ksize=5)
magnitude = cv2.addWeighted(dx, 0.5, dy, 0.5, 0)
self.magnitude = abs(magnitude)
self.phase = cv2.phase(dx, dy, angleInDegrees=True)
# calculate bins for every cell
def bins(self, cell_mag, cell_phase):
bins = np.zeros(self.bin_num)
for i in range(self.cell_size_v):
for j in range(self.cell_size_h):
quotient = int(cell_phase[i][j] / self.bin_unit)
mod = cell_phase[i][j] % self.bin_unit
if quotient == self.bin_num:
quotient = self.bin_num-1
bins[quotient] += cell_mag[i][j] * (1 - mod / self.bin_unit)
if quotient == self.bin_num - 1:
bins[0] += cell_mag[i][j] * (mod / self.bin_unit)
else:
bins[quotient + 1] += cell_mag[i][j] * (mod / self.bin_unit)
return bins
# get cell vectors(3D)
def cell_feature(self):
self.cell_vector = np.zeros([int(self.img_v / self.cell_size_v), int(self.img_h / self.cell_size_h),
self.bin_num])
for i in range(self.cell_vector.shape[0]):
for j in range(self.cell_vector.shape[1]):
cell_mag = self.magnitude[self.cell_size_v * i: self.cell_size_v * (i + 1),
self.cell_size_h * j: self.cell_size_h * (j + 1)]
cell_phase = self.phase[self.cell_size_v * i: self.cell_size_v * (i + 1),
self.cell_size_h * j: self.cell_size_h * (j + 1)]
self.cell_vector[i][j] = self.bins(cell_mag, cell_phase)
# normalize block vector
def norm_block(self, vector):
sum = 0
for i in vector:
sum += i ** 2
num = math.sqrt(sum)
if num != 0:
vector /= num
return vector
# get hog vector
def hog_feature(self):
self.cell_feature()
hog_vector = []
block_num_h = int((self.img_h - self.block_size_h) / self.block_stride_h + 1)
block_num_v = int((self.img_v - self.block_size_v) / self.block_stride_v + 1)
bczRate_h = int(self.block_size_h / self.cell_size_h) # block cell size rate
bczRate_v = int(self.block_size_v / self.cell_size_v)
bcsRate_h = int(self.block_stride_h / self.cell_size_h) # block cell stride rate
bcsRate_v = int(self.block_stride_v / self.cell_size_v)
for i in range(block_num_v):
for j in range(block_num_h):
block_vector = self.cell_vector[bcsRate_v * i : bcsRate_v * i + bczRate_v,
bcsRate_h * j : bcsRate_h * j + bczRate_h]
block_vector = block_vector.reshape(bczRate_h * bczRate_v * self.bin_num, 1)
block_vector = self.norm_block(block_vector)
hog_vector.extend(block_vector)
return hog_vector
# get image
def hog_image(self):
hog_img = np.zeros([self.img_v, self.img_h])
max = np.array(self.cell_vector).max()
halfcell_h = self.cell_size_h / 2
halfcell_v = self.cell_size_v / 2
for i in range(int(self.img_v / self.cell_size_v)):
for j in range(int(self.img_h / self.cell_size_h)):
cell_grad = self.cell_vector[i][j] / max
angle = 0
for magnitude in cell_grad:
angle_radian = math.radians(angle)
x1 = int(i * self.cell_size_h + halfcell_h + magnitude * halfcell_h * math.cos(angle_radian))
y1 = int(j * self.cell_size_v + halfcell_v + magnitude * halfcell_v * math.sin(angle_radian))
x2 = int(i * self.cell_size_h + halfcell_h - magnitude * halfcell_h * math.cos(angle_radian))
y2 = int(j * self.cell_size_v + halfcell_v - magnitude * halfcell_v * math.sin(angle_radian))
cv2.line(hog_img, (y1, x1), (y2, x2), int(255 * math.sqrt(magnitude)))
angle += self.bin_unit
return hog_img
def hog_extract(self):
self.gradient()
hog_vector = self.hog_feature()
hog_img = self.hog_image()
return hog_vector, hog_img
class extract_feature:
def __init__(self, dir_save_feature, fn_feature, data, block_size, block_stride, cell_size, bin_num):
self.dir_save_feature = dir_save_feature
self.fn_feature = fn_feature
self.data = data
self.block_size = block_size
self.block_stride = block_stride
self.cell_size = cell_size
self.bin_num = bin_num
def make_dir(self,path):
isExist = os.path.exists(path)
if not isExist:
os.makedirs(path)
def HoG_output_vector(self):
if os.path.exists(self.dir_save_feature + self.fn_feature + ".npy"):
self.np_final_vector = np.load(self.dir_save_feature + self.fn_feature + ".npy")
print(" ========= " + self.fn_feature + "Feature File is already exist =========")
else:
final_vector = []
#final_image = []
for frame in range(np.shape(self.data)[0]): # for frame in range(label.shape[0]):
if frame % 3000 == 0:
print(" >>>> Extracte Frame No." + str(frame))
temp_img = self.data[frame, :, :]
hog = HoG(temp_img, self.block_size, self.block_stride, self.cell_size, self.bin_num)
vector, image = hog.hog_extract() # vector特征，image图像
np_vector = np.array(vector)
final_vector.append(np_vector)
self.np_final_vector = np.array(final_vector) # (46233, 121, 36)
# final_image = np.append(final_image, image)
self.make_dir(self.dir_save_feature)
np.save(self.dir_save_feature + self.fn_feature + ".npy", self.np_final_vector)
return self.np_final_vector