def runMSRCP(imgPath):
with open('config.json', 'r') as f:
config = json.load(f)
img = cv2.imread(imgPath, 1)
img_msrcp = retinex.MSRCP(img, config['sigma_list'], config['low_clip'],
config['high_clip'])
m3F.imshow(img_msrcp, "MSRCP")
return img_msrcp
def retinex_shadow_removal(image, retinex_type):
with open('config.json', 'r') as f:
config = json.load(f)
if retinex_type == 'automatedMSRCR':
return retinex.automatedMSRCR(image, config['sigma_list'])
elif retinex_type == 'MSRCRP':
return retinex.MSRCP(image, config['sigma_list'], config['low_clip'],
config['high_clip'])
elif retinex_type == 'MSRCR':
return retinex.MSRCR(image, config['sigma_list'], config['G'],
config['b'], config['alpha'], config['beta'],
config['low_clip'], config['high_clip'])
else:
return image
def runMSRCP(imgPath):
with open('config.json', 'r') as f:
config = json.load(f)
img = cv2.imread(imgPath, 1)
shape = img.shape
img_msrcp = retinex.MSRCP(img, config['sigma_list'], config['low_clip'],
config['high_clip'])
img_msrcp = cv2.cvtColor(img_msrcp, cv2.COLOR_RGB2BGR)
plt.title("MSRCP")
plt.imshow(img_msrcp)
plt.show()
return img_msrcp
with open('config.json', 'r') as f:
config = json.load(f)
# 遍历读入的图像
for img_name in img_list:
if img_name == '.gitkeep':
continue
# 读入图像
img = cv2.imread(os.path.join(data_path, img_name))
img_histogram = coloredHistoEqual(img)
img_gamma = adjust_gamma(img)
# 多尺度Retinex带色彩恢复
img_msrcr = retinex.MSRCR(img, config['sigma_list'], config['G'],
config['b'], config['alpha'], config['beta'],
config['low_clip'], config['high_clip'])
img_amsrcr = retinex.automatedMSRCR(img, config['sigma_list'])
# MSR with chromaticity preservation
img_msrcp = retinex.MSRCP(img, config['sigma_list'], config['low_clip'],
config['high_clip'])
# 实验结果
shape = img.shape
cv2.imshow('Image', img)
cv2.imshow('retinex', img_msrcr)
cv2.imshow('Histogram equalization', img_histogram)
cv2.imshow('Gamma Correction', img_gamma)
cv2.imshow('Automated retinex', img_amsrcr)
cv2.imshow('MSRCP', img_msrcp)
cv2.waitKey()
def LOMO(img, config):
'''
Extract LOMO features from image
Input:
img: numpy array of shape (height, width, channels): image to extract features from
config: dict: dictionary of configuration parameters
Output:
lomo: numpy array: LOMO feature vector (length depending on image size and configuration parameters)
'''
# load relevant configuration parameters from config dictionary
sigma_list = config['retinex']['sigma_list']
G = config['retinex']['G']
b = config['retinex']['b']
alpha = config['retinex']['alpha']
beta = config['retinex']['beta']
low_clip = config['retinex']['low_clip']
high_clip = config['retinex']['high_clip']
R_list = config['lomo']['R_list']
tau = config['lomo']['tau']
hsv_bin_size = config['lomo']['hsv_bin_size']
block_size = config['lomo']['block_size']
block_step = config['lomo']['block_step']
# perform retinex transformation on image
img_retinex = retinex.MSRCP(img, sigma_list, low_clip, high_clip)
# initialize SILTP feature vector and HSV color feature vector
siltp_feat = np.array([])
hsv_feat = np.array([])
# process original image and two downsampled versions by pooling
for pool in range(3):
# determine the number of subwindows to process row- and column-wise
row_num = (img.shape[0] - (block_size - block_step)) / block_step
col_num = (img.shape[1] - (block_size - block_step)) / block_step
# iterate over all subwindows
for row in range(int(row_num)):
for col in range(int(col_num)):
# extract subwindow from image
img_block = img[row * block_step:row * block_step + block_size,
col * block_step:col * block_step + block_size]
# initialize array of multiple SILTP histograms
siltp_hist = np.array([])
# multiple SILTP histograms are computed for different radii given by configuration
for R in R_list:
# compute SILTP features for every pixel in the subwindow
siltp4 = siltp.SILTP4(img_block, R, tau)
# get bins and frequency for each bin
unique, count = np.unique(siltp4, return_counts=True)
# initialize histogram with zeros
siltp_hist_r = np.zeros([3**4])
# fill histogram
for u, c in zip(unique, count):
siltp_hist_r[u] = c
# concatenate SILTP for all radii
siltp_hist = np.concatenate([siltp_hist, siltp_hist_r], 0)
# extract subwindow from retinex image
img_block = img_retinex[row * block_step:row * block_step +
block_size,
col * block_step:col * block_step +
block_size]
# transform retinex image to HSV color space for color feature extraction
img_hsv = cv2.cvtColor(img_block, cv2.COLOR_BGR2HSV)
# build color histogram based on HSV image
hsv_hist = channel_histogram.jointHistogram(
img_hsv, [0, 255], hsv_bin_size)
# compute maximal occurrence over all subwindows of the same row
if col == 0:
siltp_feat_col = siltp_hist
hsv_feat_col = hsv_hist
else:
# maximal occurrence for SILTP histogram
siltp_feat_col = np.maximum(siltp_feat_col, siltp_hist)
# maximal occurrence for HSV histogram
hsv_feat_col = np.maximum(hsv_feat_col, hsv_hist)
# concatenate the histograms over all rows
siltp_feat = np.concatenate([siltp_feat, siltp_feat_col], 0)
hsv_feat = np.concatenate([hsv_feat, hsv_feat_col], 0)
# apply pooling and perform feature extraction with downsampled image
img = averagePooling(img)
img_retinex = averagePooling(img_retinex)
# log transform
siltp_feat = np.log(siltp_feat + 1.0)
# normalize to unit length
siltp_feat[:siltp_feat.shape[0] // 2] /= np.linalg.norm(
siltp_feat[:siltp_feat.shape[0] // 2])
siltp_feat[siltp_feat.shape[0] // 2:] /= np.linalg.norm(
siltp_feat[siltp_feat.shape[0] // 2:])
# log transform
hsv_feat = np.log(hsv_feat + 1.0)
# normalize to unit length
hsv_feat /= np.linalg.norm(hsv_feat)
# concatenate all features to final LOMO descriptor
lomo = np.concatenate([siltp_feat, hsv_feat], 0)
return lomo
def LOMO(img, config):
sigma_list = config['retinex']['sigma_list']
G = config['retinex']['G']
b = config['retinex']['b']
alpha = config['retinex']['alpha']
beta = config['retinex']['beta']
low_clip = config['retinex']['low_clip']
high_clip = config['retinex']['high_clip']
R_list = config['lomo']['R_list']
tau = config['lomo']['tau']
hsv_bin_size = config['lomo']['hsv_bin_size']
block_size_col = config['lomo']['block_size_col']
block_step_col = config['lomo']['block_step_col']
block_size_row = config['lomo']['block_size_row']
block_step_row = config['lomo']['block_step_row']
img_retinex = retinex.MSRCP(img, sigma_list, low_clip, high_clip)
siltp_feat = np.array([])
hsv_feat = np.array([])
for pool in range(3):
row_num = int((img.shape[0] - (block_size_row - block_step_row)) /
block_step_row)
col_num = int((img.shape[1] - (block_size_col - block_step_col)) /
block_step_col)
for row in range(row_num):
for col in range(col_num):
img_block = img[row * block_step_row:row * block_step_row +
block_size_row,
col * block_step_col:col * block_step_col +
block_size_col]
siltp_hist = np.array([])
for R in R_list:
siltp4 = siltp.SILTP4(img_block.astype(np.float32), R, tau)
unique, count = np.unique(siltp4, return_counts=True)
siltp_hist_r = np.zeros([3**4])
for u, c in zip(unique, count):
siltp_hist_r[u] = c
siltp_hist = np.concatenate([siltp_hist, siltp_hist_r], 0)
img_block = img_retinex[row *
block_step_row:row * block_step_row +
block_size_row, col *
block_step_col:col * block_step_col +
block_size_col]
img_hsv = cv2.cvtColor(img_block.astype(np.float32),
cv2.COLOR_BGR2HSV)
hsv_hist = channel_histogram.jointHistogram(
img_hsv, [0, 255], hsv_bin_size)
if col == 0:
siltp_feat_col = siltp_hist
hsv_feat_col = hsv_hist
else:
siltp_feat_col = np.maximum(siltp_feat_col, siltp_hist)
hsv_feat_col = np.maximum(hsv_feat_col, hsv_hist)
siltp_feat = np.concatenate([siltp_feat, siltp_feat_col], 0)
hsv_feat = np.concatenate([hsv_feat, hsv_feat_col], 0)
img = averagePooling(img)
img_retinex = averagePooling(img_retinex)
siltp_feat = np.log(siltp_feat + 1.0)
if int(siltp_feat.shape[0]) % 2 == 1:
print("ATTENTION")
siltp_feat[:int(siltp_feat.shape[0] / 2)] /= np.linalg.norm(
siltp_feat[:int(siltp_feat.shape[0] / 2)])
siltp_feat[int(siltp_feat.shape[0] / 2):] /= np.linalg.norm(
siltp_feat[int(siltp_feat.shape[0] / 2):])
hsv_feat = np.log(hsv_feat + 1.0)
hsv_feat /= np.linalg.norm(hsv_feat)
lomo = np.concatenate([siltp_feat, hsv_feat], 0)
return lomo
