def test_blind():
im = plt.imread('original/lena.bmp')
gray = images.make_gray(im)
gray = np.float64(gray)
#h=convolves.gaussian(13,15,15)
h = convolves.motion_blur(20, 30)
#con=convolves.convolution2(gray, h)
con = convolves.convolution2(gray, h)
plt.imsave(fname='l_r_blind/real_h.bmp',
arr=np.uint8(images.correct_image(h * 255)),
cmap='gray')
filt, new_h = filters.lucy_richardson_blind_deconvolution0_1(
images.make0to1(con), 50, 1, images.make0to1(gray))
plt.imsave(fname='l_r_blind/new_lena.bmp',
arr=np.uint8(images.correct_image(filt)),
cmap='gray')
plt.figure()
plt.subplot(1, 5, 1)
plt.imshow(gray, cmap='gray')
plt.title('Original')
plt.subplot(1, 5, 2)
plt.imshow(np.uint8(images.correct_image(con)), cmap='gray')
plt.title('Convoluton')
plt.subplot(1, 5, 3)
plt.imshow(np.uint8(images.correct_image(filt)), cmap='gray')
plt.title('L-R blind\nn=20, m=10')
plt.subplot(1, 5, 4)
plt.imshow(h, cmap='gray')
plt.title('PSF\ngaussian\nsigma=15, 13x13')
plt.subplot(1, 5, 5)
plt.imshow(new_h, cmap='gray')
plt.title('new PSF')
plt.show()
def test_l_r_graph():
im = plb.imread("original/my_bike.jpg")
#h=convolves.motion_blur(20,30)
gray = images.make_gray(im)
h = convolves.gaussian(10, 15, 15)
con = convolves.convolution(gray, h)
#con, noise=convolves.add_normal_noise(con, 0, 1)
start = time.time()
print 'go'
filt = filters.lucy_richardson_deconvolution(con,
h,
eps=0,
original=gray,
N=82)
print 'end'
end = time.time()
f = filt
#print im.shape, filt[:512,:512,:3].shape
#print images.compare_images_rgb(im, f)
print end - start
#np.set_printoptions(threshold=np.nan)
print con
plt.figure()
plt.subplot(1, 4, 1)
plt.imshow(gray, cmap='gray')
plt.title('original')
plt.subplot(1, 4, 2)
plt.imshow(np.uint8(images.correct_image(con)), cmap='gray')
plt.title('Gaussian blur\nsigma=10, size=15x15\nnoise~N(0,1)')
#plt.title('Motion blur\nlen=20, ang=30\nnoise~N(0,1)')
plt.subplot(1, 4, 3)
plt.imshow(np.uint8(images.correct_image(filt)), cmap='gray')
#plt.title('Tikhonov regulerization\ngamma=1e-02')
plt.title('Lucy-Richardson deconvolution\neps=5000')
plt.subplot(1, 4, 4)
plt.imshow(h, cmap='gray')
plt.title('PSF')
plt.show()
def test_pir():
im = plt.imread('original/lena.bmp')
gray = images.make_gray(im)
gray = np.float64(gray)
h = convolves.gaussian(13, 15, 15)
#h=convolves.motion_blur(20,30)
#print 'real_h =',h
#con=convolves.convolution2(gray, h)
con = convolves.convolution(gray, h)
plt.imsave(fname='l_r_blind/real_h.bmp',
arr=np.uint8(images.correct_image(h * 255)),
cmap='gray')
filt, new_h = filters.lucy_richardson_blind_deconvolution_pir(
con, 5, 1, 1, 501, 'gaussian', original=gray)
#filt=images.correct_image(filt*255)
plt.imsave(fname='l_r_blind/new_lena.bmp',
arr=np.uint8(images.correct_image(filt)),
cmap='gray')
plt.figure()
plt.subplot(1, 5, 1)
plt.imshow(gray, cmap='gray')
plt.title('Original')
plt.subplot(1, 5, 2)
plt.imshow(np.uint8(images.correct_image(con)), cmap='gray')
plt.title('Convoluton')
plt.subplot(1, 5, 3)
plt.imshow(np.uint8(images.correct_image(filt)), cmap='gray')
plt.title('L-R blind\nn=20, m=10')
plt.subplot(1, 5, 4)
plt.imshow(h, cmap='gray')
plt.title('PSF\nmotion blur\nlen=20 ang=30')
plt.subplot(1, 5, 5)
plt.imshow(new_h, cmap='gray')
plt.title('new PSF')
#print images.compare_images(gray, con[:g])
#print images.compare_images(gray, filt)
plt.show()
def lucy_richardson_blind_deconvolution_pir(g, n, m,d, max_psf_size=0, init_h_mode='gaussian', original=None):
# g - blurred image
# m, n - count of interation in RL-method
# max_psf_size - How long must be PSF
# init_h_mode - How initialize PSF:
# 'gaussian' - gaussian PSF size = 3x3, sigma =
# 'horizontal' - motion blur size 3x3, ang=0
# 'vertical' - motion blir size 3x3 ang = 90
# 'wow' - init with g
plt.imsave(fname='l_r_blind/g.bmp', arr=np.uint8(g), cmap='gray')
if max_psf_size==0:
max_psf_size = min(g.shape[0], g.shape[1])
#
# init h:
#
if init_h_mode=='gaussian':
h = conv.gaussian(1,3,3)
elif init_h_mode == 'horizontal':
h = np.array([[0,0,0],
[1,1,1],
[0,0,0]], dtype=float)/3.
elif init_h_mode == 'vertical':
h =np.array([[0,1,0],
[0,1,0],
[0,1,0]], dtype=float)/3.
elif init_h_mode == 'wow':
mini_g=smisc.imresize(g, (3,3))
h = (1. / np.sum(mini_g) ** 2) * conv.correlation2(mini_g, mini_g)
h /= np.sum(h)
else:
h = conv.random_psf(3,3)
print h
s=3
err=[]
while (s<=max_psf_size):
print 'size =',s
mini_g=smisc.imresize(g, (s,s))
f = lucy_richardson_deconvolution(mini_g, h, 20000)
if (s!=3):
h=smisc.imresize(h, (s,s))
for k in range (d):
if(k!=0):
f = lucy_richardson_deconvolution(mini_g, h, 20000)
for i in range(n):
#f_prev = np.copy(f)
#Correct h
h_prev=np.copy(h)
for k in range(m):
p = mini_g / (conv.convolution2(f, h))
flr = np.fliplr(np.flipud(f))
h = conv.convolution2(p, flr)*h
#h /= np.sum(h)
#print 'h =',h
for k in range(m):
p = mini_g / (conv.convolution2(h_prev, f))
hlr = np.fliplr(np.flipud(h_prev))
f = conv.convolution2(p, hlr) * f
print (float(i + 1) / n) * 100, ' %'
name = 'l_r_blind/new_lena' +str(s)+'_'+ str(i) + '.bmp'
h_name = 'l_r_blind/h_'+str(s)+'_' + str(i) + '.bmp'
plt.imsave(fname=name, arr=np.uint8(images.correct_image(f)), cmap='gray')
plt.imsave(fname=h_name, arr=np.uint8(images.correct_image(h*255)), cmap='gray')
s=int(s*math.sqrt(2))
#
if(original!=None):
good_f = lucy_richardson_deconvolution(g, h, 20000)
print good_f.shape
dx = (good_f.shape[0]-original.shape[0])//2
dy = (good_f.shape[1]-original.shape[1])//2
err.append(images.compare_images(good_f[dx:original.shape[0]+dx,
dy:original.shape[1] + dy],
original))
err=np.array(err)
plt.figure()
plt.plot(np.arange(err.size), err)
#plt.imshow(h,cmap='gray')
plt.show()
f = lucy_richardson_deconvolution(g, h, 5000)
return f, h
def lucy_richardson_blind_deconvolution0_1(g, n, m, original):
#init h
print 'First error',images.compare_images(g, original)
plt.imsave(fname='l_r_blind/g.bmp', arr=np.uint8(images.make0to255(g)), cmap='gray')
#h1=conv.random_psf(3,3)
# h1[1,:3]=1/3.
##h=np.zeros(g.shape, dtype=float)
#h[255:258, 255:258]=h1
h = (1. / np.sum(g) ** 2) * conv.correlation2(g, g)
h/=np.sum(h)
#h=conv.random_psf(g.shape[0], g.shape[1])
#h/=np.sum(h)
#h = conv.correlation2(g, g)
# print np.sum(h)
#h/=np.sum(h)
plt.imsave(fname='l_r_blind/init_h.bmp', arr=np.uint8(images.make0to255(h)), cmap='gray')
print 'h sum=', np.sum(h)
#h[255:258, 255:258]=h1
f=np.copy(g)
print 'blind l-r'
print '0 %'
errors=[]
for i in range(n):
f_prev=np.copy(f)
#print 'h:'
#print '-- 0 %'
for k in range(m):
p = g / (conv.convolution2(f, h))
flr=np.fliplr(np.flipud(f))
h=conv.convolution2(p,flr)*h
h/=np.sum(f)
#h/=np.sum(h)
#h = (1. / np.sum(f)) * (h * conv.correlation2(f, p))
# p=g/(sg.convolve2d(f,h, mode='same'))
# h=(1./np.sum(f))*(h*sg.correlate2d(f,p,mode='same'))
#print '--', float(k+1) / m * 100, '%'
# f=(1./np.sum(h))*(f*sg.correlate2d(h,p,mode='same'))
# print 'f:'
# print '-- 0 %'
h /= np.sum(h)
for k in range(m):
p = g / (conv.convolution2(f, h))
hlr = np.fliplr(np.flipud(h))
f = conv.convolution2(p, hlr) * f
# f = (1. / np.sum(h)) * (f * conv.correlation2(h, p))
# p=g/(sg.convolve2d(f,h, mode='same'))
# print '--', float(k+1) / m * 100, '%'
#images.check_image(f)
print (float(i+1) / n) * 100, ' %'
name='l_r_blind/new_lena'+str(i)+'.bmp'
h_name='l_r_blind/h_'+str(i)+'.bmp'
plt.imsave(fname=name, arr=np.uint8(images.make0to255(f)), cmap='gray')
#print 't and f comp', images.compare_images(temp, f)
plt.imsave(fname=h_name, arr=np.uint8(images.make0to255(h)), cmap='gray')
images.check_image(images.correct_image(f))
error = images.compare_images(images.correct_image(f), original)
errors.append(error)
print 'err =',error
errors=np.array(errors)
plt.figure()
plt.plot(np.arange(errors.size), errors)
plt.xlabel('Steps')
plt.ylabel('Dif btw original and f_k')
plt.show()
return f,h