-
Notifications
You must be signed in to change notification settings - Fork 7
/
predict.py
157 lines (141 loc) · 6.47 KB
/
predict.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
import argparse
import os
import numpy as np
from PIL import Image, ImageDraw
import torch
from torchvision import transforms
import cfg
from label import point_inside_of_quad
from model_VGG import advancedEAST
from preprocess import resize_image
from nms import nms
def sigmoid(x):
"""`y = 1 / (1 + exp(-x))`"""
return 1 / (1 + np.exp(-x))
def cut_text_line(geo, scale_ratio_w, scale_ratio_h, im_array, img_path, s):
geo /= [scale_ratio_w, scale_ratio_h]
p_min = np.amin(geo, axis=0)
p_max = np.amax(geo, axis=0)
min_xy = p_min.astype(int)
max_xy = p_max.astype(int) + 2
sub_im_arr = im_array[min_xy[1]:max_xy[1], min_xy[0]:max_xy[0], :].copy()
for m in range(min_xy[1], max_xy[1]):
for n in range(min_xy[0], max_xy[0]):
if not point_inside_of_quad(n, m, geo, p_min, p_max):
sub_im_arr[m - min_xy[1], n - min_xy[0], :] = 255
sub_im = Image.fromarray(sub_im_arr.astype('uint8')).convert('RGB')
sub_im.save(img_path + '_subim%d.jpg' % s)
def predict(east_detect, img_path, pixel_threshold, quiet=False):
img = Image.open(img_path) # 为PIL图像对象,默认RGB
d_wight, d_height = resize_image(img, cfg.max_predict_img_size)
transform = transforms.Compose([
transforms.Resize((d_wight, d_height), interpolation=2),
transforms.ToTensor()
])
x = transform(img)
x = torch.unsqueeze(x, 0) # 增加一个维度
y = east_detect(x)
y = torch.squeeze(y, 0) # 减少一个维度
print(y.shape)
y = y.detach().numpy() # 7*64*64
if y.shape[0] == 7:
y = y.transpose((1, 2, 0)) # CHW->HWC
y[:, :, :3] = sigmoid(y[:, :, :3])
cond = np.greater_equal(y[:, :, 0], pixel_threshold)
activation_pixels = np.where(cond)
quad_scores, quad_after_nms = nms(y, activation_pixels)
with Image.open(img_path) as im:
im_array = np.array(im.convert('RGB')) # 图片转为numpy数组
d_wight, d_height = resize_image(im, cfg.max_predict_img_size)
scale_ratio_w = d_wight / im.width
scale_ratio_h = d_height / im.height
im = im.resize((d_wight, d_height), Image.NEAREST).convert('RGB')
quad_im = im.copy()
draw = ImageDraw.Draw(im)
for i, j in zip(activation_pixels[0], activation_pixels[1]):
px = (j + 0.5) * cfg.pixel_size
py = (i + 0.5) * cfg.pixel_size
line_width, line_color = 1, 'red'
if y[i, j, 1] >= cfg.side_vertex_pixel_threshold:
if y[i, j, 2] < cfg.trunc_threshold:
line_width, line_color = 2, 'yellow'
elif y[i, j, 2] >= 1 - cfg.trunc_threshold:
line_width, line_color = 2, 'green'
draw.line([(px - 0.5 * cfg.pixel_size, py - 0.5 * cfg.pixel_size),
(px + 0.5 * cfg.pixel_size, py - 0.5 * cfg.pixel_size),
(px + 0.5 * cfg.pixel_size, py + 0.5 * cfg.pixel_size),
(px - 0.5 * cfg.pixel_size, py + 0.5 * cfg.pixel_size),
(px - 0.5 * cfg.pixel_size, py - 0.5 * cfg.pixel_size)],
width=line_width, fill=line_color)
im.save(img_path + '_act.jpg')
quad_draw = ImageDraw.Draw(quad_im)
txt_items = []
for score, geo, s in zip(quad_scores, quad_after_nms,
range(len(quad_scores))):
if np.amin(score) > 0:
quad_draw.line([tuple(geo[0]),
tuple(geo[1]),
tuple(geo[2]),
tuple(geo[3]),
tuple(geo[0])], width=2, fill='red')
if cfg.predict_cut_text_line:
cut_text_line(geo, scale_ratio_w, scale_ratio_h, im_array,
img_path, s)
rescaled_geo = geo / [scale_ratio_w, scale_ratio_h] # (N, 4, 2)标签坐标
rescaled_geo_list = np.reshape(rescaled_geo, (8,)).tolist()
txt_item = ','.join(map(str, rescaled_geo_list))
txt_items.append(txt_item + '\n')
elif not quiet:
print('quad invalid with vertex num less then 4.')
quad_im.save(img_path + '_predict.jpg')
if cfg.predict_write2txt and len(txt_items) > 0:
with open(img_path[:-4] + '.txt', 'w') as f_txt:
f_txt.writelines(txt_items)
def predict_txt(east_detect, img_path, txt_path, pixel_threshold, quiet=False):
img = Image.open(img_path) # 为PIL图像对象,默认RGB
d_wight, d_height = resize_image(img, cfg.max_predict_img_size)
scale_ratio_w = d_wight / img.width
scale_ratio_h = d_height / img.height
transform = transforms.Compose([
transforms.Resize((d_wight, d_height), interpolation=2),
transforms.ToTensor()
])
x = transform(img)
x = torch.unsqueeze(x, 0) # 增加一个维度
y = east_detect(x)
y = torch.squeeze(y, 0) # 减少一个维度
print(y.shape)
y = y.detach().numpy() # 7*64*64
if y.shape[0] == 7:
y = y.transpose((1, 2, 0)) # CHW->HWC
y[:, :, :3] = sigmoid(y[:, :, :3])
cond = np.greater_equal(y[:, :, 0], pixel_threshold)
activation_pixels = np.where(cond)
quad_scores, quad_after_nms = nms(y, activation_pixels)
txt_items = []
for score, geo in zip(quad_scores, quad_after_nms):
if np.amin(score) > 0:
rescaled_geo = geo / [scale_ratio_w, scale_ratio_h]
rescaled_geo_list = np.reshape(rescaled_geo, (8,)).tolist()
txt_item = ','.join(map(str, rescaled_geo_list))
txt_items.append(txt_item + '\n')
elif not quiet:
print('quad invalid with vertex num less then 4.')
if cfg.predict_write2txt and len(txt_items) > 0:
with open(txt_path, 'w') as f_txt:
f_txt.writelines(txt_items)
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument('--path', '-p', default='demo/TB1AelCLXXXXXcFXFXXunYpLFXX.jpg', help='image path')
parser.add_argument('--threshold', '-t', default=cfg.pixel_threshold, help='pixel activation threshold')
return parser.parse_args()
if __name__ == '__main__':
args = parse_args()
img_path = args.path
if not os.path.exists('demo'):
os.makedirs('./demo', exist_ok=True)
threshold = float(args.threshold)
print(img_path, threshold)
east = advancedEAST()
east.load_state_dict(torch.load(cfg.saved_model_weights_file_path))
predict(east, img_path, threshold)