/
imageutils.py
182 lines (167 loc) · 7.21 KB
/
imageutils.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
import numpy
import os.path
import subprocess
import PIL.Image
import PIL.ImageDraw
import PIL.ImageFont
import six
def read(ipath,dtype=numpy.float32):
'''
Returns a H x W x 3 RGB image in the range of [0,1].
'''
img=PIL.Image.open(ipath)
if img.mode!='RGB':
img=img.convert('RGB')
return numpy.asarray(img,dtype=dtype)/255.0
def read_shape(ipath,dtype=numpy.float32):
'''
Returns the shape (height, width, 3) of the image. This is often much
faster than reading the image.
'''
img=PIL.Image.open(ipath)
return (img.height, img.width, 3)
def write(opath,I,**kwargs):
'''
Given a H x W x 3 RGB image it is clipped to the range [0,1] and
written to an 8-bit image file.
'''
img=PIL.Image.fromarray((I*255).clip(0,255).astype(numpy.uint8))
ext=os.path.splitext(opath)[1]
if ext=='.jpg':
quality=kwargs['quality'] if 'quality' in kwargs else 95
img.save(opath,quality=quality,optimize=True)
elif ext=='.png':
img.save(opath)
else:
# I do not want to save unknown extensions because there is no
# expectation that the default save options are reasonable.
raise ValueError('Unknown image extension ({})'.format(ext))
def write_bytes(I,ext,**kwargs):
'''
Given a H x W x 3 RGB image it is clipped to the range [0,1] and
returned as a byte buffer. ext is the file extension (.jpg or .png).
'''
img=PIL.Image.fromarray((I*255).clip(0,255).astype(numpy.uint8))
f=six.StringIO()
if ext=='.jpg':
quality=kwargs['quality'] if 'quality' in kwargs else 95
img.save(f,format='JPEG',quality=quality,optimize=True)
elif ext=='.png':
img.save(f,format='PNG')
else:
# I do not want to save unknown extensions because there is no
# expectation that the default save options are reasonable.
raise ValueError('Unknown image extension ({})'.format(ext))
return f.getvalue()
def resize(I,shape):
image=PIL.Image.fromarray((I*255).clip(0,255).astype(numpy.uint8))
if shape[0]<I.shape[0] and shape[1]<I.shape[1]:
image=image.resize((shape[1],shape[0]),PIL.Image.BICUBIC)
else:
image=image.resize((shape[1],shape[0]),PIL.Image.LANCZOS)
return numpy.asarray(image,dtype=I.dtype)/255.0
def scale(I,scale_factor):
return resize(I,(int(round(I.shape[0]*scale_factor)),int(round(I.shape[1]*scale_factor))))
def resize_to_fit(I,shape):
scale_factor=min(float(shape[0])/I.shape[0],float(shape[1])/I.shape[1])
return scale(I,scale_factor)
def resize_to_cover(I,shape):
scale_factor=max(float(shape[0])/I.shape[0],float(shape[1])/I.shape[1])
return scale(I,scale_factor)
def montage(M,sep=0,canvas_value=0):
# row X col X H X W X C
assert M.ndim==5
canvas=numpy.ones((M.shape[0]*M.shape[2]+(M.shape[0]-1)*sep,M.shape[1]*M.shape[3]+(M.shape[1]-1)*sep,M.shape[4]),dtype=M.dtype)*canvas_value
for i in range(M.shape[0]):
for j in range(M.shape[1]):
canvas[i*(M.shape[2]+sep):i*(M.shape[2]+sep)+M.shape[2],j*(M.shape[3]+sep):j*(M.shape[3]+sep)+M.shape[3]]=M[i,j]
return canvas
def montage_fixed_width(S):
'''
S is a list of N rows. Each row is a list of H x W x 3 images.
Returns a rectangular image where each row is the same width.
'''
Hmean=[sum(y.shape[0] for y in x)//len(x) for x in S] # mean height of unscaled rows
scale=[[Hmean[i]/float(y.shape[0]) for y in x] for i,x in enumerate(S)] # scale equalizes image height in each row
Wsum=[sum(y.shape[1]*scale[i][j] for j,y in enumerate(x)) for i,x in enumerate(S)] # width of height-scaled rows
Wmean=int(sum(Wsum)//len(S)) # mean width
scale2=[[Wmean/float(Wsum[i])*scale[i][j] for j,y in enumerate(x)] for i,x in enumerate(S)] # scale equalizes image height in each row and row width
Hmean2=[int(sum(y.shape[0]*scale2[i][j] for j,y in enumerate(x))//len(x)) for i,x in enumerate(S)] # mean height of unscaled rows
shape=[[(Hmean2[i],int(y.shape[1]*scale2[i][j]),3) for j,y in enumerate(x)] for i,x in enumerate(S)] # new height and idth for each image
Wsum2=[sum(shape[i][j][1] for j,y in enumerate(x)) for i,x in enumerate(S)] # width of height-scaled rows
H=sum(Hmean2)
result=numpy.empty((H,Wmean,3),dtype=S[0][0].dtype)
def resize(a,shape):
if a.dtype==numpy.uint8:
img=PIL.Image.fromarray(a)
else:
img=PIL.Image.fromarray((a*255).astype(numpy.uint8))
if shape[1]<img.size[0] or shape[0]<img.size[1]:
img=img.resize((shape[1],shape[0]),PIL.Image.LANCZOS)
else:
img=img.resize((shape[1],shape[0]),PIL.Image.BILINEAR)
if a.dtype==numpy.uint8:
result=numpy.array(img)
else:
result=(numpy.array(img)/255.0).astype(a.dtype)
return result
idx0=0
for i,x in enumerate(S):
idx1=0
for j,y in enumerate(x):
if j==len(x)-1:
# absorb rounding errors into last image of each row
result[idx0:idx0+shape[i][j][0],idx1:]=resize(S[i][j],(shape[i][j][0],result.shape[1]-idx1))
else:
result[idx0:idx0+shape[i][j][0],idx1:idx1+shape[i][j][1]]=resize(S[i][j],shape[i][j])
idx1+=shape[i][j][1]
idx0+=shape[i][0][0]
return result
try:
def render_text(s,size,font=PIL.ImageFont.truetype(subprocess.check_output(['fc-match','-f','%{file}','Droid Serif']),80)):
'''
Returns rendered text as a H x W x 3 numpy array in the range [0,1]. The
image will be the exact size given. If one dimension is None then the
image will be sized to fit the given dimension.
'''
image=PIL.Image.new('RGB',(5,5),(255,255,255))
draw=PIL.ImageDraw.Draw(image)
image=PIL.Image.new('RGB',draw.textsize(s,font=font),(255,255,255))
draw=PIL.ImageDraw.Draw(image)
draw.text((0,0),s,(0,0,0),font=font)
if size[0]==None:
size=list(size)
size[0]=int(round(image.size[1]*size[1]/float(image.size[0])))
elif size[1]==None:
size=list(size)
size[1]=int(round(image.size[0]*size[0]/float(image.size[1])))
image=image.resize((size[1],size[0]),PIL.Image.LANCZOS)
I=numpy.array(image)/255.0
return I
except:
def render_text(*args,**kwargs):
'''
Returns rendered text as a H x W x 3 numpy array in the range [0,1]. The
image will be the exact size given. If one dimension is None then the
image will be sized to fit the given dimension.
'''
raise RuntimeError('Could not define render_text (probably because fc-match failed).')
def concatenate(X,axis,canvas_value=0,gravity=(-1)):
'''
Given a sequence of images, concatenate them along the given axis,
expanding the other axes as needed. If gravity is zero then the original
data will be centered in the output domain. Negative or positive gravity
will cause it to be flush with the lower or upper bound, respectively.
'''
outshape=[sum(x.shape[i] for x in X) if i==axis else max(x.shape[i] for x in X) for i in range(X[0].ndim)]
Y=[]
for x in X:
newshape=list(outshape)
newshape[axis]=x.shape[axis]
if gravity>0:
Y.append(numpy.pad(x,[(newshape[i]-x.shape[i],0) for i in range(x.ndim)],'constant',constant_values=canvas_value))
elif gravity==0:
Y.append(numpy.pad(x,[((newshape[i]-x.shape[i])//2,(newshape[i]-x.shape[i])-(newshape[i]-x.shape[i])//2) for i in range(x.ndim)],'constant',constant_values=canvas_value))
else:
Y.append(numpy.pad(x,[(0,newshape[i]-x.shape[i]) for i in range(x.ndim)],'constant',constant_values=canvas_value))
return numpy.concatenate(Y,axis=axis)