def equal_contrast(gauss):
#step1
alfa = 0.1
ec = gauss
ec0 = cv2.absdiff(ec, 0)
ec0 = cv2.pow(ec0, alfa)
ec0 = cv2.mean(ec0)
ec0 = ec0[0]
den = np.power(ec0, 1.0 / alfa)
ec1 = cv2.divide(ec, den)
#step2
tal = 10.0
ec2 = cv2.absdiff(ec1, 0)
ec2 = cv2.min(tal, ec2)
ec2 = cv2.pow(ec2, alfa)
ec2 = cv2.mean(ec2)
ec2 = ec2[0]
den = np.power(ec2, 1.0 / alfa)
ec3 = cv2.divide(ec1, den)
ec3 = cv2.normalize(ec3.astype('float32'), None, 0.0, 1.0, cv2.NORM_MINMAX)
#step3
ele = cv2.divide(ec3, tal)
exp1 = cv2.exp(ele)
exp2 = cv2.exp(-ele)
num = cv2.subtract(exp1, exp2)
den = cv2.add(exp1, exp2)
tanh = den
tanh = cv2.divide(num, den, tanh, tal)
ec4 = tanh
ec4 = cv2.normalize(ec4.astype('float32'), None, -1.0, 1.3,
cv2.NORM_MINMAX)
return ec4
def tantriggs(image):
# Convert to float
image = np.float32(image)
image = cv2.pow(image, GAMMA)
image = difference_of_gaussian(image)
# mean 1
tmp = cv2.pow(cv2.absdiff(image, 0), ALPHA)
mean = cv2.mean(tmp)[0]
image = cv2.divide(image, cv2.pow(mean, 1.0 / ALPHA))
# mean 2
tmp = cv2.pow(cv2.min(cv2.absdiff(image, 0), TAU), ALPHA)
mean = cv2.mean(tmp)[0]
image = cv2.divide(image, cv2.pow(mean, 1.0 / ALPHA))
# tanh
exp_x = cv2.exp(cv2.divide(image, TAU))
exp_negx = cv2.exp(cv2.divide(-image, TAU))
image = cv2.divide(cv2.subtract(exp_x, exp_negx), cv2.add(exp_x, exp_negx))
image = cv2.multiply(image, TAU)
image = cv2.normalize(image, None, 0, 255, cv2.NORM_MINMAX, cv2.CV_8UC1)
return image
def tantriggs(image):
# Convert to float
image = np.float32(image)
image = cv2.pow(image, GAMMA)
image = difference_of_gaussian(image)
# mean 1
tmp = cv2.pow(cv2.absdiff(image, 0), ALPHA)
mean = cv2.mean(tmp)[0]
image = cv2.divide(image, cv2.pow(mean, 1.0/ALPHA))
# mean 2
tmp = cv2.pow(cv2.min(cv2.absdiff(image, 0), TAU), ALPHA)
mean = cv2.mean(tmp)[0]
image = cv2.divide(image, cv2.pow(mean, 1.0/ALPHA))
# tanh
exp_x = cv2.exp(cv2.divide(image, TAU))
exp_negx = cv2.exp(cv2.divide(-image, TAU))
image = cv2.divide(cv2.subtract(exp_x, exp_negx), cv2.add(exp_x, exp_negx))
image = cv2.multiply(image, TAU)
image = cv2.normalize(image, None, 0, 255, cv2.NORM_MINMAX, cv2.CV_8UC1)
return image
def ShowHomomorphicFilter(imgSrc):
# imgSrc = cv.resize(imgSrc, (int(imgSrc.shape[1]/4), int(imgSrc.shape[0]/4)))
imgLnSrc = imgSrc
# 先把范围控制下，不然0值被log以后会出无限值
cv.normalize(imgLnSrc, imgLnSrc, 1, 255, cv.NORM_MINMAX)
imgLnSrc = np.float64(imgLnSrc)
cv.log(imgLnSrc, imgLnSrc)
b, g, r = cv.split(imgLnSrc)
H = ImageHandler.CreateHomomorphicFilterTemplate(
imgLnSrc.shape[0] * 2, imgLnSrc.shape[1] * 2)
b = ImageHandler.HandleFFTPerChannel(b, H)
g = ImageHandler.HandleFFTPerChannel(g, H)
r = ImageHandler.HandleFFTPerChannel(r, H)
merged = cv.merge((b, g, r))
imgOut = merged[0:imgSrc.shape[0], 0:imgSrc.shape[1],
0:imgSrc.shape[2]]
# 两次归一，是因为出来的大小太大了，exp会出无限值
cv.normalize(imgOut, imgOut, 0, 1, cv.NORM_MINMAX)
cv.exp(imgOut, imgOut)
cv.normalize(imgOut, imgOut, 0, 1, cv.NORM_MINMAX)
cv.cvtColor(imgSrc, cv.COLOR_BGR2RGB, imgSrc)
cv.cvtColor(imgOut, cv.COLOR_BGR2RGB, imgOut)
plt.figure(1), plt.imshow(imgSrc)
plt.figure(2), plt.imshow(imgOut)
plt.show()
def noniternorm(img):
b,g,r = cv2.split(img)
b = np.float32(b)
g = np.float32(g)
r = np.float32(r)
log_b = cv2.log(b)
log_g = cv2.log(g)
log_r = cv2.log(r)
b = cv2.exp(log_b - cv2.mean(log_b)[0])
g = cv2.exp(log_g - cv2.mean(log_g)[0])
r = cv2.exp(log_r - cv2.mean(log_r)[0])
b = cv2.normalize(b, 0, 255, cv2.NORM_MINMAX)*255
g = cv2.normalize(g, 0, 255, cv2.NORM_MINMAX)*255
r = cv2.normalize(r, 0, 255, cv2.NORM_MINMAX)*255
return cv2.merge((np.uint8(b),np.uint8(g),np.uint8(r)))
def sigmoid(x, height, width):
zeroes = cv2.UMat(np.zeros((height, width)))
ones = cv2.UMat(np.ones((height, width)))
subbed = cv2.subtract(zeroes, x, dtype=cv2.CV_32F)
added = cv2.add(ones, cv2.exp(subbed), dtype=cv2.CV_32F)
return cv2.divide(ones, added, dtype=cv2.CV_32F)
def saliency_feature(img):
img_orig = img
img = cv2.resize(img, (64, 64))
img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
# h = cv2.getOptimalDFTSize(img.shape[0])
# w = cv2.getOptimalDFTSize(img.shape[1])
# print "Resizing (%d, %d) to (%d, %d)" % (img.shape[0], img.shape[1], h, w)
# h = (h - img.shape[0])/2.0
# w = (w - img.shape[1])/2.0
# img = cv2.copyMakeBorder(img, int(math.floor(h)), int(math.ceil(h)), int(math.floor(w)), int(math.ceil(w)), cv2.BORDER_CONSTANT, value=0)
dft = cv2.dft(np.float32(img), flags=cv2.DFT_COMPLEX_OUTPUT)
A, P = cv2.cartToPolar(dft[:,:,0], dft[:,:,1])
L = cv2.log(A)
h_n = (1./3**2)*np.ones((3,3))
R = L - cv2.filter2D(L, -1, h_n)
S = cv2.GaussianBlur(cv2.idft(np.dstack(cv2.polarToCart(cv2.exp(R), P)), flags=cv2.DFT_REAL_OUTPUT)**2, (0,0), 8)
S = cv2.resize(cv2.normalize(S, None, 0, 1, cv2.NORM_MINMAX), (img_orig.shape[1],img_orig.shape[0]))
# cv2.namedWindow('tmp1', cv2.WINDOW_NORMAL)
# cv2.imshow('tmp1', img_orig)
# cv2.namedWindow('tmp', cv2.WINDOW_NORMAL)
# cv2.imshow('tmp', S)
# cv2.waitKey()
return S
def saliency_feature(img):
img_orig = img
img = cv2.resize(img, (64, 64))
img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
# h = cv2.getOptimalDFTSize(img.shape[0])
# w = cv2.getOptimalDFTSize(img.shape[1])
# print "Resizing (%d, %d) to (%d, %d)" % (img.shape[0], img.shape[1], h, w)
# h = (h - img.shape[0])/2.0
# w = (w - img.shape[1])/2.0
# img = cv2.copyMakeBorder(img, int(math.floor(h)), int(math.ceil(h)), int(math.floor(w)), int(math.ceil(w)), cv2.BORDER_CONSTANT, value=0)
dft = cv2.dft(np.float32(img), flags=cv2.DFT_COMPLEX_OUTPUT)
A, P = cv2.cartToPolar(dft[:, :, 0], dft[:, :, 1])
L = cv2.log(A)
h_n = (1. / 3**2) * np.ones((3, 3))
R = L - cv2.filter2D(L, -1, h_n)
S = cv2.GaussianBlur(
cv2.idft(np.dstack(cv2.polarToCart(cv2.exp(R), P)),
flags=cv2.DFT_REAL_OUTPUT)**2, (0, 0), 8)
S = cv2.resize(cv2.normalize(S, None, 0, 1, cv2.NORM_MINMAX),
(img_orig.shape[1], img_orig.shape[0]))
# cv2.namedWindow('tmp1', cv2.WINDOW_NORMAL)
# cv2.imshow('tmp1', img_orig)
# cv2.namedWindow('tmp', cv2.WINDOW_NORMAL)
# cv2.imshow('tmp', S)
# cv2.waitKey()
return S
def gamma(rng, img, gamma_range):
gamma = rng_between(rng, gamma_range[0], gamma_range[1])
k = 1.0 / gamma
img = cv2.exp(k * cv2.log(img.astype('float32') + 1e-15))
f = math.pow(255.0, 1 - k)
img = img * f
img = cv2.add(img, np.zeros_like(img), dtype=0) # clip
return img
def noniternorm(img):
b, g, r = cv2.split(img)
b = np.float32(b)
g = np.float32(g)
r = np.float32(r)
log_b = cv2.log(b)
log_g = cv2.log(g)
log_r = cv2.log(r)
b = cv2.exp(log_b - cv2.mean(log_b)[0])
g = cv2.exp(log_g - cv2.mean(log_g)[0])
r = cv2.exp(log_r - cv2.mean(log_r)[0])
b = cv2.normalize(b, 0, 255, cv2.NORM_MINMAX) * 255
g = cv2.normalize(g, 0, 255, cv2.NORM_MINMAX) * 255
r = cv2.normalize(r, 0, 255, cv2.NORM_MINMAX) * 255
b = b.clip(max=255)
g = g.clip(max=255)
r = r.clip(max=255)
return cv2.merge((np.uint8(b), np.uint8(g), np.uint8(r)))
def dFFT2(img):
"""
Calculated the discrete Fourier
Transfrom of a 2D image
:param img: image to be transformed
:return: complex valued result
"""
N=len(img)
if N==1:
return img
even=dFFT2([img[k] for k in range(0,N,2)])
odd= dFFT2([img[k] for k in range(1,N,2)])
M=N/2
l=[ even[k] + cv2.exp(-2j*3.14*k/N)*odd[k] for k in range(M) ]
r=[ even[k] - cv2.exp(-2j*3.14*k/N)*odd[k] for k in range(M) ]
return l+r
def split_r_part(hsv_v, nkernelSize, mean_illumination, ADJUST_DELTA=10):
v_blur = cv2.GaussianBlur(hsv_v, (nkernelSize, nkernelSize), 0)
v_log = cv2.log(hsv_v)
v_blur_log = cv2.log(v_blur)
r_part_log = v_log - v_blur_log
r_part = cv2.exp(r_part_log)
r_part = cv2.convertScaleAbs(r_part, alpha=mean_illumination)
r_part_32F = r_part.astype(np.float32)
r_part_32F -= ADJUST_DELTA
return r_part_32F
def __init__(self, image, parent=None):
super(ResamplingWidget, self).__init__(parent)
filt_size = 3
em_radius = 2
em_stddev = 5
em_error = 0.01
max_iter = 20
gray = cv.cvtColor(image, cv.COLOR_BGR2GRAY).astype(np.float32)
minimum, maximum, _, _ = cv.minMaxLoc(gray)
kernel = np.array([[1, 2, 1], [2, 4, 2], [1, 2, 1]], np.float32)
kernel /= (filt_size + 1) * (filt_size + 1)
alpha0 = np.ones((2 * em_radius + 1, 2 * em_radius + 1), np.float32)
alpha0[em_radius, em_radius] = 0
alpha0 /= alpha0.size - 1
alpha1 = np.zeros((2 * em_radius + 1, 2 * em_radius + 1), np.float32)
gray0 = np.ravel(gray[em_radius:-em_radius, em_radius:-em_radius])
block_rows = gray.shape[0] - 2 * em_radius
block_cols = gray.shape[1] - 2 * em_radius
resamp = np.array([])
for i in range(-em_radius, em_radius + 1):
for j in range(-em_radius, em_radius + 1):
if i == 0 and j == 0:
continue
block = gray[i + em_radius:i + em_radius + block_rows,
j + em_radius:j + em_radius + block_cols]
block = np.reshape(block, (block.size, 1))
resamp = block if resamp.size == 0 else np.hstack(
(resamp, block))
i = 0
sigma = em_stddev
c1 = 1 / (sigma * np.sqrt(2 * np.pi))
c2 = 2 * sigma**2
p0 = 1 / (maximum - minimum)
while cv.norm(alpha0, alpha1) > em_error and i < max_iter:
filt = cv.filter2D(gray, cv.CV_32F, alpha1)
resid0 = cv.absdiff(gray, filt)
resid = resid0[em_radius:resid0.shape[0] - em_radius,
em_radius:resid0.shape[0] - em_radius]
resid = cv.pow(cv.filter2D(resid, cv.CV_32F, kernel), 2)
cond = c1 * (1 / cv.exp(resid / c2))
post = cond / (cond + p0)
sigma = np.sqrt(np.sum(post * resid) / np.sum(post)) / 2
alpha0 = np.copy(alpha1)
weights = np.reshape(post, (post.size, 1))
# deriv = np.multiply()
k = 0
def _merge(self, images, times):
"""
:Description: use images, times, and CRF to merge HDRI
:param images: image list
:param times: times list
:return hdr_img: HDRI(lux_img)
"""
assert isinstance(images, list), 'images should be list'
assert isinstance(times, list), 'times should be list'
assert len(images) == len(
times), "images length should be same as times"
weights = self.__intensity_weight_256x_.copy()
n_img = len(images)
n_chn = images[0].shape[2]
response_256x1x3 = self.__camera_response_256x1x3.copy()
log_response = np.log(response_256x1x3)
log_time = np.log(times)
# log_hdr_img channel list
hdr_chn_list = [0, 0, 0]
img_avr_w_sum = np.zeros(images[0].shape[:2], dtype="float32")
for i in xrange(n_img):
src_chn_list = cv2.split(images[i])
img_avr_w = np.zeros(images[0].shape[:2], dtype="float32")
for cn in xrange(n_chn):
img_cn_w = cv2.LUT(src_chn_list[cn], weights)
img_avr_w += img_cn_w
#第n张图3个通道的平均权值图像
img_avr_w /= n_chn
#一张图的log_response(log(lum))
response_img = cv2.LUT(images[i], log_response)
response_chn_list = cv2.split(response_img)
for chn in xrange(n_chn):
#img_avr_w:图片的平均通道权值 response_chn_list[chn]:通道的log_response log_time[i]:图片的log_time.
hdr_chn_list[chn] += cv2.multiply(
img_avr_w, response_chn_list[chn] - log_time[i])
#全部图的平均权值的和
img_avr_w_sum += img_avr_w
#全部图的平均权值的和的倒数
img_avr_w_sum = 1.0 / img_avr_w_sum
for cn in xrange(n_chn):
hdr_chn_list[cn] = cv2.multiply(hdr_chn_list[cn], img_avr_w_sum)
log_hdr_img = cv2.merge(hdr_chn_list)
#this is lux, 为什么和官方的数值有数量级的差别。
hdr_img = cv2.exp(log_hdr_img)
return hdr_img
def process(self, images, times):
"""
:Description: combine factors weight to merge HDRI and tonermap LDRI
:param images: images list
:param times: times list
:return ldr_img: LDRI
"""
assert isinstance(images, list), 'images should be list'
assert isinstance(times, list), 'times should be list'
assert len(images) == len(times), "images length should be same as times"
start = time.time()
gamma = self.__gamma
contrast = self.__contrast
saturation = self.__saturation
sigma_space = self.__sigma_space
sigma_color = self.__sigma_color
hdr_img = self._merge(images, times)
hdr_img_2d = hdr_img.reshape(1024, 1280*3)
minval, maxvalue, _, _ = cv2.minMaxLoc(hdr_img_2d)
img = (hdr_img - minval) / (maxvalue - minval)
img = img.clip(1.0e-4)
img = cv2.pow(img, 1.0 / gamma)
gray_img = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)
log_img = np.log(gray_img)
map_img = cv2.bilateralFilter(log_img, -1, sigma_color, sigma_space)
minval, maxval, _, _ = cv2.minMaxLoc(map_img)
scale = contrast / (maxval - minval)
map_img = cv2.exp(map_img * (scale - 1.0) + log_img)
img = self._mapLuminance(img, gray_img, map_img, saturation)
img = cv2.pow(img, 1.0 / gamma)
#no problem!!
img = img.clip(None, 1.0)
img = img * 255
ldr_img = img.astype("uint8")
end = time.time()
print "spend time %f" % (end-start)
return ldr_img
def multi_scale_retinex(image, ksize=0, sigma=9):
I = image.copy()
I = cv2.convertScaleAbs(np.float32(I), I, alpha=1.0, beta=1.0)
I = cv2.log(np.float32(I), I)
l = cv2.GaussianBlur(image, (ksize, ksize), sigma)
l = cv2.convertScaleAbs(np.float32(l), l, alpha=1.0, beta=1.0)
l = cv2.log(np.float32(l), l)
result = cv2.subtract(I, l, l)
result = cv2.exp(result, result)
frames = cv2.split(result)
for i in range(3):
frames[i] = cv2.normalize(frames[i],
frames[i],
alpha=0,
beta=255,
norm_type=cv2.NORM_MINMAX)
frames[i] = cv2.convertScaleAbs(frames[i],
frames[i],
alpha=1 / 3,
beta=0.0)
return color_correction(cv2.merge(frames))
def sovle_c_w(self, contrast, wellexp, images, shape, i, is_gray):
if is_gray:
contrast[i] = cv2.Laplacian(images[i][:, :, 0], cv2.CV_32F)
else:
gray = cv2.cvtColor(images[i], cv2.COLOR_RGB2GRAY)
contrast[i] = cv2.Laplacian(gray, cv2.CV_32F)
contrast[i] = np.abs(contrast[i])
if self.wcon != 1:
contrast[i] = cv2.pow(contrast[i], self.wcon)
wellexp[i] = np.ones(shape, dtype="float32")
if self.wexp != 0:
splitted = [
images[i][:, :, 0], images[i][:, :, 1], images[i][:, :, 2]
]
for img_cn in splitted:
expo = cv2.subtract(img_cn, 0.5, dtype=cv2.CV_32F)
expo = cv2.pow(expo, 2.0)
expo = -expo / 0.08
# larger '0.08' only make 'cv2.exp(expo)' nearest "1"
expo = cv2.exp(expo)
wellexp[i] *= expo
wellexp[i] = cv2.pow(wellexp[i], self.wexp)
return 1.0 / (2.0 * pow(x, 2))
print(b.shape, type(b[0, 0]))
print(b[0])
print(h.shape, type(h[0, 0]))
print(h[0])
while (1):
cv2.imshow('image', img)
k = cv2.waitKey(1) & 0xFF
if k == 27:
break
hv = cv2.getTrackbarPos('H', 'image')
sv = cv2.getTrackbarPos('S', 'image')
sigma = cv2.getTrackbarPos('Sigma', 'image')
hv = hv - 90
temp0 = h - ht
temp1 = 0.5 * (temp0 - hv)
temp2 = 0.1 * (s - sv)
cv2.imshow('b image', temp1)
cv2.imshow('g image', temp2)
temp0 = cv2.pow(temp1, 2)
temp1 = cv2.pow(temp2, 2)
temp2 = temp0 + temp1
temp0 = -(temp2 * cal_sigma(sigma))
temp1 = cv2.exp(temp0)
temp2 = temp1 * 255.0
cv2.imshow('r image', temp2)
cv2.destroyAllWindows()
# from (Book) OpenCV-Python으로 배우는 영상 처리 및 응용
import numpy as np, cv2
# numpy array 생성 예시
v1 = np.array([1, 2, 3], np.float32) # 1차원 리스트로 생성- 행벡터
v2 = np.array([[1], [2], [3]], np.float32) # 2차원 리스트로(3행, 1열) - 행벡터
v3 = np.array([[1, 2, 3]], np.float32) # 2차원 리스트로(1행, 3열) - 일반 행렬
# OpenCV 산술 연산 함수는 numpy array만 가능함
v1_exp = cv2.exp(v1) # 벡터에 대한
v2_exp = cv2.exp(v2) # 행렬에 대한 지수 계산
v3_exp = cv2.exp(v3) # 행렬에 대한 지수 계산
log = cv2.log(v1) # 로그 계산
sqrt= cv2.sqrt(v2) # 제곱근 계산
pow = cv2.pow(v3, 3) # 3의 거듭제곱 계산
# 결과 출력
print("[v1] 형태: %s 원소: %s" % ( v1.shape, v1))
print("[v2] 형태: %s 원소:\n%s" % ( v2.shape, v2))
print("[v3] 형태: %s 원소: %s" % ( v3.shape, v3))
print()
# 행렬 정보 출력 - OpenCV 결과는 행렬로 반환됨 - 행벡터는 열벡터로 반환됨
print("[v1_exp] 자료형: %s 형태: %s" % ( type(v1_exp), v1_exp.shape)) # 행벡터 인수의 결과
print("[v2_exp] 자료형: %s 형태: %s" % ( type(v2_exp), v2_exp.shape)) # 행벡터 인수의 결과
print("[v3_exp] 자료형: %s 형태: %s" % ( type(v3_exp), v3_exp.shape)) # 행벡터 인수의 결과
print()
# 열벡터를 1 행에 출력하는 예시 - 행벡터로 변환
print("[log] =", log.T)
print("[sqrt] =", np.ravel(sqrt))
1. ravel()
2. flatten()
3. arr.T
'''
import numpy as np
import cv2
from numpy.core.fromnumeric import ndim, size
v1 = np.array([1, 2, 3], np.float32) # 3, 행렬 (dim = 1)
v2 = np.array([[1], [2], [3]],
np.float32) # 3 x 1 행렬 (dim = 2) / print(ndim(v2)) == 2
v3 = np.array([[1, 2, 3]], np.float32) # 1 x 3 행령
v_exp = cv2.exp(v1)
m_exp = cv2.exp(v2)
q_exp = cv2.exp(v3)
print(v_exp.shape)
print(m_exp.shape)
print(q_exp.shape)
v_log = cv2.log(v1)
m_sqrt = cv2.sqrt(v2)
q_pow = cv2.pow(v3, 3)
print(v_log.shape)
print(m_sqrt.shape)
print(q_pow.shape)
def process(self, images, times, samples=70, random=False):
"""
:Description: to calibrate CRF curve
:param images: image list
:param times: time list
:param samples: samples point count
:param random: whether samples random
:return: response_256x1x3, CRF array
"""
assert isinstance(images, list), 'images should be list'
assert isinstance(times, list), 'times should be list'
assert len(images) == len(
times), "images length should be same as times"
LDR_SIZE = self.__LDR_SIZE
w = self.__intensity_weight_256x_.copy()
gamma = self.__gamma
images = np.array(images, dtype="uint8")
times = np.array(times, dtype="float32")
n_img = len(images)
n_chn = images[0].shape[2]
img_channel_list = []
for i in xrange(n_chn):
tmp = []
for j in xrange(n_img):
img_channel = cv2.split(images[j])[i]
tmp.append(img_channel)
img_channel_list.append(tmp)
img_shape = images[0].shape
img_cols = img_shape[1]
img_rows = img_shape[0]
sample_points_list = []
#set random situation.
if random == True:
for i in xrange(samples):
r = np.random.randint(0, img_rows)
c = np.random.randint(0, img_cols)
sample_points_list.append((r, c))
if random == False:
x_points = int(np.sqrt(samples * (img_cols) / img_rows))
y_points = samples / x_points
n_samples = x_points * y_points
step_x = img_cols / x_points
step_y = img_rows / y_points
r = step_x / 2
for j in xrange(y_points):
rr = r + j * step_y
c = step_y / 2
for i in xrange(x_points):
cc = c + i * step_x
sample_points_list.append((rr, cc))
#svd solve response curve.
response_list = []
for z in xrange(n_chn):
eq = 0
A = np.zeros(
(n_samples * n_img + LDR_SIZE + 1, LDR_SIZE + n_samples),
dtype="float32")
B = np.zeros((A.shape[0]), dtype="float32")
for i in xrange(n_samples):
r = sample_points_list[i][0]
c = sample_points_list[i][1]
for j in xrange(n_img):
val = img_channel_list[z][j][r, c]
A[eq, val] = w[val]
A[eq, LDR_SIZE + i] = -w[val]
B[eq] = w[val] * np.log(times[j])
eq += 1
#F(128)曝光量对数设0, 也就是曝光量为单位1, 不关事
A[eq, LDR_SIZE / 2] = 1
eq += 1
for i in range(0, 254):
A[eq, i] = gamma * w[i]
A[eq, i + 1] = -2 * gamma * w[i]
A[eq, i + 2] = gamma * w[i]
eq += 1
_, response = cv2.solve(A, B, flags=cv2.DECOMP_SVD)
# just from ln(lum) convert to lum.
response = cv2.exp(response)
response_256x1 = response[:256]
response_list.append(response_256x1)
self.camera_response_256x1x3 = cv2.merge(response_list)
#need return 256x1x3 nparray.
return self.camera_response_256x1x3
def process(self, images):
"""
:rtype: object
:Description: combine factors weight to fusion HDRI-alternate
:param images: img list
:return fusion_img: fusion_img
"""
assert isinstance(images, list), 'images should be list'
n_img = len(images)
#camera images both are [b, g, r] 3channels
# n_chn = images[0].shape[2]
n_chn = 1
rows, cols = images[0].shape[0], images[0].shape[1]
shape = (rows, cols)
#initial the weights[]
weights = []
for i in xrange(n_img):
tmp = np.zeros(shape, dtype="float32")
weights.append(tmp)
weight_sum = np.zeros(shape, dtype="float32")
#solve each image weight and solve weight_sum
for i in xrange(len(images)):
#normalize img make convergence speedup
img = images[i] / 255.0
##############################
# img = images[i] / 65536.0
img = img.astype("float32")
#convert to gray
if n_chn == 3:
gray = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)
else:
gray = img
#solve contrast, ddepth: cv2.CV_32F
contrast = cv2.Laplacian(gray, cv2.CV_32F)
contrast = abs(contrast)
#
mean = np.zeros(shape, dtype="float32")
splitted = cv2.split(images[i])
for img_cn in splitted:
mean += img_cn
mean /= n_chn
#solve saturation
saturation = np.zeros(shape, dtype="float32")
for img_cn in splitted:
deviation = img_cn - mean
deviation = cv2.pow(deviation, 2.0)
saturation += deviation
saturation = cv2.sqrt(saturation)
wellexp = np.ones(shape, dtype="float32")
for img_cn in splitted:
expo = cv2.subtract(img_cn, 0.5, dtype=cv2.CV_32F)
expo = cv2.pow(expo, 2.0)
expo = -expo / 0.08
#larger '0.08' only make 'cv2.exp(expo)' nearest "1"
expo = cv2.exp(expo)
wellexp = cv2.multiply(wellexp, expo)
# pow respective ratio
contrast = cv2.pow(contrast, self.wcon)
saturation = cv2.pow(saturation, self.wsat)
wellexp = cv2.pow(wellexp, self.wexp)
weights[i] = contrast
if n_chn == 3:
weights[i] = cv2.multiply(weights[i], saturation)
weights[i] = cv2.multiply(weights[i], wellexp) + 1e-12
weight_sum += weights[i]
maxlevel = int(np.log(min(rows, cols)) / np.log(2))
#(maxlevel+1) images, following to solve the final pyramid.
res_pyr = [0] * (maxlevel + 1)
for i in xrange(len(images)):
img_pyr = [0] * (maxlevel + 1)
weight_pyr = [0] * (maxlevel + 1)
###############################
img = images[i] / 255.0
# img = images[i] / 65535.0
img = img.astype("float32")
img_pyr[0] = img
weights[i] /= weight_sum
weight_pyr[0] = weights[i]
# following: buildPyramid(img, img_pyr, maxlevel)
# buildPyramid(weights[i], weight_pyr, maxlevel)
#todo inspection it
for lvl in xrange(maxlevel):
img_pyr[lvl + 1] = cv2.pyrDown(img_pyr[lvl])
for lvl in xrange(maxlevel):
#size = width, height
size = img_pyr[lvl].shape[:2][::-1]
up = cv2.pyrUp(img_pyr[lvl + 1], dstsize=size)
img_pyr[lvl] -= up
for lvl in xrange(maxlevel):
weight_pyr[lvl + 1] = cv2.pyrDown(weight_pyr[lvl])
for lvl in xrange(maxlevel + 1):
splitted = cv2.split(img_pyr[lvl])
splitted2 = []
for img_pyr_cn in splitted:
tmp = cv2.multiply(img_pyr_cn, weight_pyr[lvl])
splitted2.append(tmp)
cv2.merge(splitted2, img_pyr[lvl])
# add 3 laplace pyr together -> res_pyr
# first image to assign res_pry[0-maxlevel]
if i == 0:
res_pyr[lvl] = img_pyr[lvl]
# latter image to assign res_pry[0-maxlevel]
else:
res_pyr[lvl] += img_pyr[lvl]
#第一层求出第0层在第一次loop就已经完成了，后面的loop没有意义
for lvl in range(maxlevel, 0, -1):
#size = width, height
size = res_pyr[lvl - 1].shape[:2][::-1]
up = cv2.pyrUp(res_pyr[lvl], dstsize=size)
res_pyr[lvl - 1] += up
dst_tmp = res_pyr[0]
dst_tmp = dst_tmp * 255
# dst_tmp = dst_tmp * 65535
fusion_img = dst_tmp.astype("uint8")
# fusion_img = dst_tmp.astype("uint16")
return fusion_img
def Retinex(input_img,filter=['Gauss'],ksize=[3],weight=[1],gstd=0,hstd=0): # gstd-空间高斯函数标准差;sstd-灰度值相似性高斯函数标准差
'''Retinex匀光算法(SSR/MSR)流程:
1.估计照度分量;2.计算反射分量;3.反射分量增强
***注意***
输出input_img必须为单通道图像,即len(input_img.shape)=2
'''
multi_scale = len(filter)
blur_img = []
# 检查输入图像维度
if(len(input_img.shape)!=2):
raise RuntimeError('The dimensionality of input image must be 2')
# 检查filter,ksize和weight三者维度是否一致
if(len(filter)!=len(ksize) or len(filter)!=len(weight)):
raise RuntimeError('The number of filter, ksize and weight are not equal')
# 检查滤波器大小和权重和
if(multi_scale==1):
if (ksize[0] % 2 != 1):
raise RuntimeError('Filter ksize must be odd')
if(weight[0]!=1):
raise RuntimeError('The weight must be 1')
else:
for i in ksize:
if(i % 2 != 1):
raise RuntimeError('Filter ksize must be odd')
if(sum(weight)!=1):
raise RuntimeError('The sum of weight must be 1')
# 计算照度分量
def filter_blur(input_img=input_img,filter=filter[0],ksize=ksize[0],gstd=gstd,hstd=hstd):
if (filter == 'Gauss'):
blur = cv2.GaussianBlur(input_img, (ksize, ksize), gstd)
elif (filter == 'Mean'):
blur = cv2.blur(input_img, (ksize, ksize))
elif (filter == 'Median'):
blur = cv2.medianBlur(input_img, ksize)
elif (filter == 'Bilateral'):
blur = cv2.bilateralFilter(input_img, ksize, gstd, hstd)
else:
raise RuntimeError('Filter type error')
return blur
if(multi_scale==1):
blur_img.append(filter_blur())
else:
for i in range(len(filter)):
blur_img.append(filter_blur(filter=filter[i],ksize=ksize[i]))
# 计算反射图像:R(x,y)=f(x,y)/f(x,y)*G(x,y) => log[R(x,y)]=log[f(x,y)]-log[f(x,y)*G(x,y)] =>R(x,y)=e^{log[f(x,y)]-log[f(x,y)*G(x,y)]}
def replace_zeroes(input_img): # 替换图像中亮度为0的像素点
min_nonzero=min(input_img[np.nonzero(input_img)]) # np.nonzero返回表征非零元素在矩阵中位置的元组,元组中前一个列表存放非零行坐标，后一个列表存放非零元素列坐标
input_img[input_img==0]=min_nonzero
return input_img
if(multi_scale==1):
dst_input=cv2.log(replace_zeroes(input_img)/255.0) # 归一化像素值以保证cv2.log正常运行
dst_blur=cv2.log(replace_zeroes(blur_img[0])/255.0)
log_R=cv2.subtract(dst_input,dst_blur)
exp_R=cv2.exp(log_R)
dst_R=cv2.normalize(exp_R,None,0,255,cv2.NORM_MINMAX) # 归一化像素值至[0,255]
output_img=cv2.convertScaleAbs(dst_R) # 转换为uint8格式
else:
intermediate_img = []
dst_input = cv2.log(replace_zeroes(input_img) / 255.0)
for i in range(len(blur_img)):
dst_blur = cv2.log(replace_zeroes(blur_img[i]) / 255.0)
log_R = cv2.subtract(dst_input, dst_blur)
intermediate_img.append(weight[i]*log_R)
exp_R = cv2.exp(sum(intermediate_img))
dst_R = cv2.normalize(exp_R, None, 0, 255, cv2.NORM_MINMAX)
output_img = cv2.convertScaleAbs(dst_R)
# 颜色增强(暂缺)
return output_img
def tanh(x):
exp_x = cv2.exp(x)
exp_x_ = cv2.exp(-x)
fx = (exp_x - exp_x_) / (exp_x + exp_x_)
return fx
def sigmoid(x):
exp_x = cv2.exp(-x)
fx = 1.0 / (1.0 + exp_x)
return fx
ec1 = cv2.divide(ec, den)
#step2
tal = 10.0
ec2 = cv2.absdiff(ec1, 0)
ec2 = cv2.min(tal, ec2)
ec2 = cv2.pow(ec2, alfa)
ec2 = cv2.mean(ec2)
ec2 = ec2[0]
den = numpy.power(ec2, 1.0 / alfa)
ec3 = cv2.divide(ec1, den)
ec3 = cv2.normalize(ec3.astype('float32'), None, 0.0, 1.0, cv2.NORM_MINMAX)
#step3
ele = cv2.divide(ec3, tal)
exp1 = cv2.exp(ele)
exp2 = cv2.exp(-ele)
num = cv2.subtract(exp1, exp2)
den = cv2.add(exp1, exp2)
tanh = den
tanh = cv2.divide(num, den, tanh, tal)
ec4 = tanh
ec4 = cv2.normalize(ec4.astype('float32'), None, -1.0, 1.3,
cv2.NORM_MINMAX)
#if foto==332 or foto==343 or foto==358 or foto==654 or foto==658 or foto==662 or foto==665 or foto==668 or foto==672 or foto==919:
# cv2.imshow("Images", numpy.hstack([rosto, corrigida_gam, gaus, ec3, ec4]))
# cv2.waitKey(0)
#Fazendo a Transformada de Fourier da imagem
def sigmoid(self, x):
subbed = cv2.subtract(self.zeroes, x, dtype=cv2.CV_32F)
added = cv2.add(self.ones, cv2.exp(subbed), dtype=cv2.CV_32F)
return cv2.divide(self.ones, added, dtype=cv2.CV_32F)
