-
Notifications
You must be signed in to change notification settings - Fork 8
/
Copy pathImageMatting.py
180 lines (134 loc) · 6.39 KB
/
ImageMatting.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
import numpy as np
import scipy as sp
import cv2
import sys, os
def Create_Laplacian_Matrix(Img, Mask, _WinSize=5, Lambda=1e-7, Options="Linear", s=1.0):
WinSize=np.array([_WinSize,_WinSize])
HalfWinSize=(WinSize-1)/2
ImgSize=Img.shape[0:2]
AllPixelInds=np.arange(0,ImgSize[0]*ImgSize[1]).reshape(ImgSize)
TotalPixelNumInWindow=WinSize[0]*WinSize[1]
Temp=Mask[HalfWinSize[0]:-HalfWinSize[0],HalfWinSize[1]:-HalfWinSize[1]]
NumTrainingPixels=len(Temp[Temp==0.5])
NumofNoneZeroEntries=NumTrainingPixels*TotalPixelNumInWindow**2
Row_indices=np.zeros(NumofNoneZeroEntries)
Col_indices=Row_indices.copy()
Coeff_values=Row_indices.copy()
from numpy import matlib
from scipy.spatial.distance import pdist, squareform
Index=0
for i in range(HalfWinSize[0],ImgSize[0]-HalfWinSize[0]):
for j in range(HalfWinSize[1],ImgSize[1]-HalfWinSize[1]):
if Mask[i,j]==0.5: ### pixel that has unknown alphas
LocalWin=Img[i-HalfWinSize[0]:i+HalfWinSize[0]+1,j-HalfWinSize[1]:j+HalfWinSize[1]+1,:].reshape((-1,Img.shape[2]))
Xi=np.ones((LocalWin.shape[0],LocalWin.shape[1]+1))
Xi[:,:-1]=LocalWin
I=np.eye(Xi.shape[0])
VAL=None
if Options=="Linear":
VAL=np.dot(Xi,Xi.transpose())
elif Options=="Gaussiankernel":
pairwise_dists = squareform(pdist(Xi, 'euclidean'))
VAL = sp.exp((-pairwise_dists**2)/(2.0*(s**2)))
F=sp.linalg.solve((VAL+Lambda*I).transpose(), VAL.transpose())
I_F=np.eye(F.shape[0])-F
Local_Lap_Coeff=np.dot(I_F, I_F.transpose())
LocalWindowInds=AllPixelInds[i-HalfWinSize[0]:i+HalfWinSize[0]+1,j-HalfWinSize[1]:j+HalfWinSize[1]+1]
Row_indices[Index:TotalPixelNumInWindow**2+Index]=matlib.repmat(LocalWindowInds.reshape((-1,1)),1,TotalPixelNumInWindow).reshape(TotalPixelNumInWindow**2)
Col_indices[Index:TotalPixelNumInWindow**2+Index]=matlib.repmat(LocalWindowInds.reshape((-1,1)).transpose(),TotalPixelNumInWindow,1).reshape(TotalPixelNumInWindow**2)
Coeff_values[Index:TotalPixelNumInWindow**2+Index]=Local_Lap_Coeff.reshape(-1)
Index=Index+TotalPixelNumInWindow**2
Lap = sp.sparse.csr_matrix((Coeff_values, (Row_indices, Col_indices)), shape=(ImgSize[0]*ImgSize[1], ImgSize[0]*ImgSize[1]))
return Lap
def Solve_for_all_Alphas_from_Known_Alphas(Lap,C,Alpha_known,Gamma=1e-7):
Originshape=Alpha_known.shape
Alpha_known=Alpha_known.reshape((-1,1))
I=sp.sparse.eye(Lap.shape[0], Lap.shape[1]).tocsr()
B=C.dot(Alpha_known)
A=Lap+C+Gamma*I
from scipy.sparse.linalg import spsolve
Alpha=spsolve(A, B).reshape(Originshape) #Alpha=A.inv()*b
return Alpha.clip(0,1)
def Image_Matting_By_Learning(Img, Trimap_mask, WinSize=5, c=1000.0, Lambda=1e-7, Options="Linear", s=1.0):
Mask=np.ones(Trimap_mask.shape)*0.5
Mask[Trimap_mask==255]=1
Mask[Trimap_mask==0]=0
KnownVsUnknown_mask=np.zeros(Mask.shape)
KnownVsUnknown_mask[Mask!=0.5]=1
Lap=Create_Laplacian_Matrix(Img, Mask, WinSize, Lambda, Options, s)
TotalPixelNum=Mask.shape[0]*Mask.shape[1]
C=c*sp.sparse.spdiags(KnownVsUnknown_mask.reshape(TotalPixelNum),0,TotalPixelNum,TotalPixelNum).tocsr()
Alpha_known=Mask.copy()
Alpha=Solve_for_all_Alphas_from_Known_Alphas(Lap,C,Alpha_known)
return Alpha
## change color image:_Img to different feature data:Img.
def Get_newformat_data(_Img, Choice):
Img_rgb=cv2.cvtColor(_Img, cv2.COLOR_BGR2RGB)/255.0
Img_lab=cv2.cvtColor(_Img, cv2.COLOR_BGR2LAB)/255.0
Img_gray=cv2.cvtColor(_Img, cv2.COLOR_BGR2GRAY)/255.0
Img=None
if Choice=="rgblab":
Img=np.zeros((_Img.shape[0],_Img.shape[1],6))
Img[:,:,:3]=Img_rgb
Img[:,:,3:]=Img_lab
if Choice=="lab":
Img=Img_lab
if Choice=="rgb":
Img=Img_rgb
if Choice=="rgblabgradient":
Sobelx = cv2.Sobel(Img_gray,cv2.CV_64F,1,0,ksize=3)
Sobely = cv2.Sobel(Img_gray,cv2.CV_64F,0,1,ksize=3)
mag=np.sqrt(np.square(Sobelx)+np.square(Sobely))
Sobelx=Sobelx/(mag+1.0)
Sobely=Sobely/(mag+1.0)
Img=np.zeros((_Img.shape[0],_Img.shape[1],9))
Img[:,:,:3]=Img_rgb
Img[:,:,3:6]=Img_lab
Img[:,:,6]=(mag+1.0)/(mag+1.0).max()
Img[:,:,7]=Sobelx
Img[:,:,8]=Sobely
return Img
##### command line: python ImageMatting.py GT01.png Trimap1 Linear rgb 3
##### or: python ImageMatting.py GT01.png Trimap1 Gaussiankernel rgblab 3 1.0
if __name__=="__main__":
### Samples input:
# Trimap_name='./data/Trimap/Trimap1/GT01.png'
# Img_name='./data/Img/GT01.png'
# Groundtruth_name='./data/Groundtruth/GT01.png'
# Saveresults_name ='./data/Results/GT01-Trimap1-Alphas-'
# # Option="Linear"
# Option="Gaussiankernel"
# Choice="rgblab"
# # Choice="lab"
# # Choice="rgb"
# # Choice="rgblabgradient"
Img_name='./data/Img/'+sys.argv[1]
Trimap_name="./data/Trimap/"+sys.argv[2]+"/"+sys.argv[1]
Groundtruth_name='./data/Groundtruth/'+sys.argv[1]
Saveresults_name='./data/Results/'+ os.path.splitext(sys.argv[1])[0]+'-'+sys.argv[2]+'-Alphas-'
Option=sys.argv[3]
Choice=sys.argv[4]
WinSize=np.int(sys.argv[5])
Std=1.0
if Option=="Gaussiankernel":
Std=np.float(sys.argv[6])
print "Current Trimap used: ", Trimap_name
print "Options: ", Option
print "Choices: ", Choice
_Img=cv2.imread(Img_name)
Total_pixels=_Img.shape[0]*_Img.shape[1]
Img=Get_newformat_data(_Img, Choice)
Trimap=cv2.imread(Trimap_name,0)
Alpha_gt=cv2.imread(Groundtruth_name,0)
if Option=="Linear":
Alpha=Image_Matting_By_Learning(Img, Trimap, WinSize=WinSize, Options=Option)
Alpha=(Alpha*255).astype(np.uint8)
cv2.imwrite(Saveresults_name+Option+"-"+Choice+".png", Alpha)
Diff=Alpha-Alpha_gt
print "RMSE:", np.sqrt(np.square(Diff).sum()/Total_pixels)
if Option=="Gaussiankernel":
Alpha=Image_Matting_By_Learning(Img, Trimap, WinSize=WinSize, Options=Option, s=Std)
Alpha=(Alpha*255).astype(np.uint8)
cv2.imwrite(Saveresults_name+Option+"-Std-"+str(Std)+"-"+Choice+".png", Alpha)
Diff=Alpha-Alpha_gt
print "Std:", Std, "\tRMSE:", np.sqrt(np.square(Diff).sum()/Total_pixels)