def skin_detect(img):
"""img: in BGR mode"""
# initialized all-white mask
skinMask = np.zeros(img.shape[0:2], img.dtype)
img_LAB = cv2.cvtColor(img, cv2.COLOR_BGR2LAB).astype('float')
img_HSV = cv2.cvtColor(img, cv2.COLOR_BGR2HSV).astype('float')
# ellipse and HSV_test
skinMask[ ellipse_test(img_LAB[...,1], img_LAB[...,2], bound=1.0, prob=0.6)
& HSV_threshold(img_HSV[...,0], img_HSV[...,1]) ] \
= 255
# relaxed ellipse test, guarenteed by skin neighborhood
skinMask[ (skinMask ==0)
& ellipse_test(img_LAB[...,1], img_LAB[...,2]
, bound=1.25, prob=0.6)
& check_neighbor(skinMask)] = 255
# filling holes:image closing on skinMask
# http://stackoverflow.com/a/10317883/2729100
_h,_w = img.shape[0:2]
_kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE,(_h/16,_w/16))
skinMask_closed = cv2.morphologyEx(skinMask,cv2.MORPH_CLOSE,_kernel)
# skinMask_closed = skinMask
cape_util.display(np.hstack([skinMask, skinMask_closed]), name='skin mask closing before after', mode='gray')
# initialization, can't remove, otherwise mask==0 area will be random
skin = 255*np.ones(img.shape, img.dtype)
skin = cv2.bitwise_and(img, img, mask=skinMask_closed)
return skin, (skinMask_closed/255).astype(bool)
def exposure_correction(I_L, I_out, I_out_side_crr, skin_masks, faces_xywh,
_eacp_lambda_):
I_out_255 = cape_util.mag(I_out, 'float')
A = np.ones(I_out_side_crr.shape)
I_out_expo_crr = I_out_side_crr.copy()
for i in range(len(faces_xywh)):
x, y, w, h = faces_xywh[i]
face = I_out_side_crr[y:y + h, x:x + w]
# each sidelit crted face(L channel)
skin_face = cape_util.mask_skin(face, skin_masks[i][1]) # SKIN on that face
cumsum = cv2.calcHist([cape_util.mag(skin_face, 'trim')] # cumsumed hist of SKIN
, [0], None, [255], [1, 256]).T.ravel().cumsum()
# visualize cumsum of hist
plt.plot(cumsum)
plt.xlim([1, 256])
plt.show()
p = np.searchsorted(cumsum, cumsum[-1] * 0.75)
if p < 120:
print 'face skin is underexposed!'
f = (120 + p) / ((2 * p) + 1e-6)
A[y:y + h, x:x + w][face > 0] = f
# >0 means every pixel *\in S*
A_after = EACP(A, I_out_255, lambda_=_eacp_lambda_)
# to visualize A before/after EACP
cape_util.display(np.hstack([A, A_after]), mode='rainbow')
I_out_expo_crr = A_after * I_out_side_crr
# to visualize exposure corrected face
cape_util.display(cape_util.mag(I_out_expo_crr, 'trim'),
name='exposure corrected L',
mode='gray')
return I_out_expo_crr
def skin_detect(img):
"""img: in BGR mode"""
# initialized all-white mask
skinMask = np.zeros(img.shape[0:2], img.dtype)
img_LAB = cv2.cvtColor(img, cv2.COLOR_BGR2LAB).astype('float')
img_HSV = cv2.cvtColor(img, cv2.COLOR_BGR2HSV).astype('float')
# ellipse and HSV_test
skinMask[ ellipse_test(img_LAB[...,1], img_LAB[...,2], bound=1.0, prob=0.6)
& HSV_threshold(img_HSV[...,0], img_HSV[...,1]) ] \
= 255
# relaxed ellipse test, guarenteed by skin neighborhood
skinMask[
(skinMask == 0)
& ellipse_test(img_LAB[..., 1], img_LAB[..., 2], bound=1.25, prob=0.6)
& check_neighbor(skinMask)] = 255
# filling holes:image closing on skinMask
# http://stackoverflow.com/a/10317883/2729100
_h, _w = img.shape[0:2]
_kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (_h / 16, _w / 16))
skinMask_closed = cv2.morphologyEx(skinMask, cv2.MORPH_CLOSE, _kernel)
# skinMask_closed = skinMask
cape_util.display(np.hstack([skinMask, skinMask_closed]),
name='skin mask closing before after',
mode='gray')
# initialization, can't remove, otherwise mask==0 area will be random
skin = 255 * np.ones(img.shape, img.dtype)
skin = cv2.bitwise_and(img, img, mask=skinMask_closed)
return skin, (skinMask_closed / 255).astype(bool)
def face_enhancement(I_origin, _eacp_lambda_=0.2):
"""
Keyword Arguments:
I -- np.uint8, (m,n,3), BGR
_eacp_lambda_ -- float, scala
"""
_I_LAB = cv2.cvtColor(I_origin, cv2.COLOR_BGR2LAB)
I = _I_LAB[..., 0]
Base, Detail = wls_filter.wlsfilter(I)
I_out = Base # float [0,1]
I_aindane = aindane.aindane(I_origin)
faces_xywh = vj_face.face_detect(I_aindane)
faces = [I_origin[y:y + h, x:x + w] for (x, y, w, h) in faces_xywh]
skin_masks = [apa_skin.skin_detect(face) for face in faces]
_any_skin = False
for _mask in skin_masks:
_any_skin |= ((_mask[1]).any())
if (faces_xywh == []) or (not _any_skin): # face not detected
print 'face or skin not detected!'
return I_origin
_I_out_faces = [I_out[y:y + h, x:x + w]
for (x, y, w, h) in faces_xywh] #float[0,1]
S = [ cape_util.mask_skin(_I_out_faces[i], skin_masks[i][1]) \
for i in range(len(_I_out_faces)) ] # float [0,1]
# to visualize detected skin and it's (unsmoothed) hist
for idx, s in enumerate(S):
cape_util.display(cape_util.mag(s),
name='detected skin of L channel',
mode='gray')
# plot original hist(rectangles form, of S). don't include 0(masked)
plt.hist(cape_util.mag(s).ravel(), 255, [1, 256])
plt.xlim([1, 256])
plt.show()
## save image data for R silverman analysis
import os
DATA_DIR = os.path.expanduser('~/Dropbox/dropcode/r/')
# save skin pixels (value neq 0.0)
np.savetxt(DATA_DIR + str(idx) + '.csv',
s[s != 0.0].ravel(),
delimiter=',')
H = [cape_util.get_smoothed_hist(cape_util.mag(s)) for s in S]
# visualize smoothed hist
for h in H:
plt.plot(h)
plt.xlim([1, 256])
plt.show()
I_out_side_crr = sidelight_correction(_I_LAB[..., 0], I_out, H, S,
faces_xywh, _eacp_lambda_)
I_out_expo_crr = exposure_correction(_I_LAB[..., 0], I_out, I_out_side_crr,
skin_masks, faces_xywh, _eacp_lambda_)
_I_LAB[..., 0] = cape_util.mag((I_out_expo_crr + 255.0 * Detail), 'trim')
I_res = cv2.cvtColor(_I_LAB, cv2.COLOR_LAB2BGR)
return I_res
def face_enhancement(I_origin, _eacp_lambda_=0.2):
"""
Keyword Arguments:
I -- np.uint8, (m,n,3), BGR
_eacp_lambda_ -- float, scala
"""
_I_LAB = cv2.cvtColor(I_origin, cv2.COLOR_BGR2LAB)
I = _I_LAB[..., 0]
Base, Detail = wls_filter.wlsfilter(I)
I_out = Base # float [0,1]
I_aindane = aindane.aindane(I_origin)
faces_xywh = vj_face.face_detect(I_aindane)
faces = [I_origin[y:y + h, x:x + w] for (x, y, w, h) in faces_xywh]
skin_masks = [apa_skin.skin_detect(face) for face in faces]
_any_skin = False
for _mask in skin_masks:
_any_skin |= ((_mask[1]).any())
if (faces_xywh == []) or (not _any_skin): # face not detected
print 'face or skin not detected!'
return I_origin
_I_out_faces = [I_out[y:y + h, x:x + w] for (x, y, w, h) in faces_xywh] #float[0,1]
S = [ cape_util.mask_skin(_I_out_faces[i], skin_masks[i][1]) \
for i in range(len(_I_out_faces)) ] # float [0,1]
# to visualize detected skin and it's (unsmoothed) hist
for idx, s in enumerate(S):
cape_util.display(cape_util.mag(s),
name='detected skin of L channel',
mode='gray')
# plot original hist(rectangles form, of S). don't include 0(masked)
plt.hist(cape_util.mag(s).ravel(), 255, [1, 256])
plt.xlim([1, 256])
plt.show()
## save image data for R silverman analysis
import os
DATA_DIR = os.path.expanduser('~/Dropbox/dropcode/r/')
# save skin pixels (value neq 0.0)
np.savetxt(DATA_DIR+str(idx)+'.csv', s[s!=0.0].ravel(), delimiter=',')
H = [cape_util.get_smoothed_hist(cape_util.mag(s)) for s in S]
# visualize smoothed hist
for h in H:
plt.plot(h)
plt.xlim([1, 256])
plt.show()
I_out_side_crr = sidelight_correction(_I_LAB[..., 0], I_out, H, S,
faces_xywh, _eacp_lambda_)
I_out_expo_crr = exposure_correction(_I_LAB[..., 0], I_out, I_out_side_crr,
skin_masks, faces_xywh, _eacp_lambda_)
_I_LAB[..., 0] = cape_util.mag((I_out_expo_crr + 255.0 * Detail), 'trim')
I_res = cv2.cvtColor(_I_LAB, cv2.COLOR_LAB2BGR)
return I_res
def exposure_correction(I_L, I_out, I_out_side_crr, skin_masks, faces_xywh,
_eacp_lambda_):
I_out_255 = cape_util.mag(I_out, 'float')
A = np.ones(I_out_side_crr.shape)
I_out_expo_crr = I_out_side_crr.copy()
for i in range(len(faces_xywh)):
x, y, w, h = faces_xywh[i]
face = I_out_side_crr[y:y + h, x:x + w]
# each sidelit crted face(L channel)
skin_face = cape_util.mask_skin(face,
skin_masks[i][1]) # SKIN on that face
cumsum = cv2.calcHist(
[cape_util.mag(skin_face, 'trim')] # cumsumed hist of SKIN
,
[0],
None,
[255],
[1, 256]).T.ravel().cumsum()
# visualize cumsum of hist
plt.plot(cumsum)
plt.xlim([1, 256])
plt.show()
p = np.searchsorted(cumsum, cumsum[-1] * 0.75)
if p < 120:
print 'face skin is underexposed!'
f = (120 + p) / ((2 * p) + 1e-6)
A[y:y + h, x:x + w][face > 0] = f
# >0 means every pixel *\in S*
A_after = EACP(A, I_out_255, lambda_=_eacp_lambda_)
# to visualize A before/after EACP
cape_util.display(np.hstack([A, A_after]), mode='rainbow')
I_out_expo_crr = A_after * I_out_side_crr
# to visualize exposure corrected face
cape_util.display(cape_util.mag(I_out_expo_crr, 'trim'),
name='exposure corrected L',
mode='gray')
return I_out_expo_crr
def sky_enhance(X, P, Sbgr
,I_origin, sky_prob_map):
"""
ARGs:
-----
X: Alpha = x[0]; C = x[1]; S=x[2]
P: sky_porb_map
Sbgr: np.array([[[Sb, Sg, Sr]]]), mean of the sky color in current image
NOTE: computing some color-like array (e.g. beta_old) will actually exceed 255,
use 'float' instead of 'uint8'
"""
Alpha = X[0]; C = X[1]; S=X[2]
f_sky_bgr = np.array([[[255,210,160]]], dtype='uint8')
f_sky = cv2.cvtColor(f_sky_bgr, cv2.COLOR_BGR2LAB) # preferred color, in CIELab
Slab = cv2.cvtColor(Sbgr, cv2.COLOR_BGR2LAB)
f_lab = 1.0*f_sky / Slab # (f_l, f_a, f_b), correction ratio
print 'f_lab: ', f_lab
beta_old = cv2.cvtColor( cape_util.safe_convert(_2to3(S) * Sbgr, np.uint8), cv2.COLOR_BGR2LAB )
kai_old = cv2.cvtColor( cape_util.safe_convert(_2to3(C), np.uint8), cv2.COLOR_BGR2LAB )
W = np.array([[[100, 0, 0]]])
# W = np.array([[[255, 128, 128]]]) # (255,255,255) in RGB
P_3 = _2to3(P)
beta_new = cv2.add(P_3*(f_lab*beta_old), (1-P_3)*beta_old)
kai_new = cv2.add(P_3*((W+kai_old)/2.0), (1-P_3)*kai_old)
_fst = _2to3(1-Alpha)*cv2.cvtColor(beta_new.astype('uint8'), cv2.COLOR_LAB2BGR)
_lst = _2to3(Alpha)*cv2.cvtColor(kai_new.astype('uint8'), cv2.COLOR_LAB2BGR)
res = cv2.add( _fst, _lst ) # use cv2.add() to avoid overflow e.g. 150+125=255, not 20
# visualize _fst, _lst for debugging
_first = I_origin.copy(); _last = I_origin.copy()
_first[ sky_prob_to_01(sky_prob_map, thresh=0.0) ] = _fst[0]
_last[ sky_prob_to_01(sky_prob_map, thresh=0.0) ] = _lst[0]
cape_util.display( np.hstack([_first, _last] ) )
return res
def sidelight_correction(I_L, I_out, H, S, faces_xywh, _eacp_lambda_):
I_out_255 = cape_util.mag(I_out, 'float')
bimodal_Fs, D, M, B = cape_util.detect_bimodal(H)
A = np.ones(I_out_255.shape)
W = np.zeros(I_out_255.shape)
for i in range(len(bimodal_Fs)):
if bimodal_Fs[i] == True:
b = B[i]
d = D[i]
m = M[i]
f = (b - d) / (m - d)
print str(i) + 'th face: d, m, b, f', d, m, b, f
x, y, w, h = faces_xywh[i]
A[y:y + h,
x:x + w][(cape_util.mag(S[i], 'float') < m)
& (cape_util.mag(S[i], 'float') > 0)] = f #<m and \in S
global W
miu = (I_out_255[A == f]).mean()
sig = 255 * 3 # manually set, 3*sig = 120 close to 255/2
W = np.exp(-(I_out_255 - miu)**2 / sig**2)
# W[A==f] = 1.
W[...] = 1 - W[...]
# W[A!=f] =1.
W[...] = 1.
# face_i_crr = EACP(I_out_255[y:y+h,x:x+w]*A, I_out_255[y:y+h,x:x+w], lambda_=_eacp_lambda_)
# I_out_255[y:y+h, x:x+w] = face_i_crr
else:
print '? bimodal not detected on', i, 'th face'
# W[...] = 1.
cape_util.display(W, mode='rainbow')
A_after = A.copy()
if (A == 1).all():
print 'A == 1 everywhere!'
else:
A_after = EACP(A, I_L, W, lambda_=_eacp_lambda_)
I_out_side_crr = A_after * I_out_255
# to visualize A before/after EACP
cape_util.display(np.hstack([A, A_after]), mode='rainbow')
# to visualize sidelight corrected result
cape_util.display(cape_util.mag(I_out_side_crr, 'trim'),
name='sidelight corrected, L',
mode='gray')
return I_out_side_crr # float [0.0,255.0]
def sidelight_correction(I_L, I_out, H, S, faces_xywh, _eacp_lambda_):
I_out_255 = cape_util.mag(I_out, 'float')
bimodal_Fs, D, M, B = cape_util.detect_bimodal(H)
A = np.ones(I_out_255.shape)
W = np.zeros(I_out_255.shape)
for i in range(len(bimodal_Fs)):
if bimodal_Fs[i] == True:
b = B[i]
d = D[i]
m = M[i]
f = (b - d) / (m - d)
print str(i) + 'th face: d, m, b, f', d, m, b, f
x, y, w, h = faces_xywh[i]
A[y:y + h, x:x + w][(cape_util.mag(S[i], 'float') < m) &
(cape_util.mag(S[i], 'float') > 0)] = f #<m and \in S
global W
miu = (I_out_255[A == f]).mean()
sig = 255 * 3 # manually set, 3*sig = 120 close to 255/2
W = np.exp(-(I_out_255 - miu) ** 2 / sig ** 2)
# W[A==f] = 1.
W[...] = 1 - W[...]
# W[A!=f] =1.
W[...] = 1.
# face_i_crr = EACP(I_out_255[y:y+h,x:x+w]*A, I_out_255[y:y+h,x:x+w], lambda_=_eacp_lambda_)
# I_out_255[y:y+h, x:x+w] = face_i_crr
else:
print '? bimodal not detected on', i, 'th face'
# W[...] = 1.
cape_util.display(W, mode='rainbow')
A_after = A.copy()
if (A == 1).all():
print 'A == 1 everywhere!'
else:
A_after = EACP(A, I_L, W, lambda_=_eacp_lambda_)
I_out_side_crr = A_after * I_out_255
# to visualize A before/after EACP
cape_util.display(np.hstack([A, A_after]), mode='rainbow')
# to visualize sidelight corrected result
cape_util.display(cape_util.mag(I_out_side_crr, 'trim'),
name='sidelight corrected, L',
mode='gray')
return I_out_side_crr # float [0.0,255.0]
def main():
I_origin = cv2.imread(IMG_DIR + 'input_teaser.png')
lambda_ = 120
I_res = face_enhancement(I_origin, lambda_)
cape_util.display(np.hstack([I_origin, I_res]))
return 0
def main():
I_origin = cv2.imread(IMG_DIR+'pic23.jpg')
res_disp,_ = sky_enhancement(I_origin)
cape_util.display(np.hstack([I_origin, res_disp]))
return 0
def _test():
I_bgr = cv2.imread(IMG_DIR + 'pic36.jpg')
I_res_bgr = ss_enhance(I_bgr)
cape_util.display(np.hstack([I_bgr, I_res_bgr]))
return
def main():
img_name = 'pic23.jpg'
I_org = cv2.imread(IMG_DIR+ img_name)
skin_prob_map = apa_skin.skin_prob_map(I_org)
lambda_ = 120
res_skin = I_org
res_skin = face_enhancement.face_enhancement(I_org, lambda_)
cape_util.display( np.hstack([I_org,res_skin]), name='res_skin' )
res_sky, sky_prob_map = sky_enhancement.sky_enhancement(res_skin)
cape_util.display( np.hstack([I_org,res_sky]), name='res_sky' )
# res_ge = global_enhace(res_sky, skin_prob_map ,sky_prob_map)
res_ge = global_enhace(res_skin, skin_prob_map ,sky_prob_map)
cape_util.display( np.hstack([I_org,res_ge]), name='res_ge' )
res_ss = ss_enhance.ss_enhance(res_ge)
cape_util.display( np.hstack([I_org,res_ss]), name='res_ss' )
res_de = detail_enhace(res_ss, skin_prob_map, sky_prob_map)
cape_util.display( np.hstack([I_org,res_de]), name='res_de' )
# res_de = res_ss
cape_util.display( np.hstack([I_org,res_de]), name='lambda_='+str(lambda_) )
DIR = './resources/results/'
cv2.imwrite( DIR+str.split(img_name, '.')[0]+'_res.png', np.hstack([I_org, res_de]) )
return 0
def _test():
I_bgr = cv2.imread(IMG_DIR+'pic36.jpg')
I_res_bgr = ss_enhance(I_bgr)
cape_util.display(np.hstack([I_bgr, I_res_bgr]))
return
def main():
I_origin = cv2.imread(IMG_DIR + 'input_teaser.png')
lambda_ = 120
I_res = face_enhancement(I_origin, lambda_)
cape_util.display(np.hstack([I_origin, I_res]))
return 0