def __init__(self, batch_size,image_size,path_size):
self.data_pair1_root = "../dataset/For_pair_IMG_Train/pair1/" # assume this one is the newest frame
self.data_pair2_root = "../dataset/For_pair_IMG_Train/pair2/" # assume this one is the historical image
self.data_mat_root = "../dataset/For_pair_IMG_Train/CostMatrix/"
self.signalroot ="../dataset/For_pair_IMG_Train/saved_stastics/"
self.read_all_flag=0
self.read_record =0
self.folder_pointer = 0
self.slice_record=1
self.batch_size = batch_size
self.img_size = Resample_size
self.path_size = Path_length
self.mat_size = Mat_size
self.img_size2 = Resample_size2
# Initialize the inout for the tainning
self.input_mat = np.zeros((batch_size,1,Mat_size,Resample_size)) #matri
self.input_path = np.zeros((batch_size,Path_length))#path
self.input_pair1 = np.zeros((batch_size,1,Resample_size2,Resample_size))#pairs
self.input_pair2 = np.zeros((batch_size,1,Resample_size2,Resample_size))
self.input_pair3 = np.zeros((batch_size,1,Resample_size2,Resample_size))
self.input_pair4 = np.zeros((batch_size,1,Resample_size2,Resample_size))
# the number isdeter by teh mat num
self.all_dir_list = os.listdir(self.data_mat_root)
self.folder_num = len(self.all_dir_list)
# create the buffer list(the skill to create the list)
self.folder_mat_list = [None]*self.folder_num
self.folder_pair1_list = [None]*self.folder_num
self.folder_pair2_list = [None]*self.folder_num
self.signal = [None]*self.folder_num
# create all the folder list and their data list
number_i = 0
# all_dir_list is subfolder list
#creat the image list point to the STASTICS TIS list
saved_stastics = MY_ANALYSIS()
#read all the folder list of mat and pairs and path
for subfold in self.all_dir_list:
#the mat list
this_folder_list = os.listdir(os.path.join(self.data_mat_root, subfold))
this_folder_list2 = [ self.data_mat_root +subfold + "/" + pointer for pointer in this_folder_list]
self.folder_mat_list[number_i] = this_folder_list2
#the pair1 list
this_folder_list = os.listdir(os.path.join(self.data_pair1_root, subfold))
this_folder_list2 = [ self.data_pair1_root +subfold + "/" + pointer for pointer in this_folder_list]
self.folder_pair1_list[number_i] = this_folder_list2
#the pair2 list
this_folder_list = os.listdir(os.path.join(self.data_pair2_root, subfold))
this_folder_list2 = [ self.data_pair2_root +subfold + "/" + pointer for pointer in this_folder_list]
self.folder_pair2_list[number_i] = this_folder_list2
#the supervision signal list
#change the dir firstly before read
saved_stastics.all_statics_dir = os.path.join(self.signalroot, subfold, 'signals.pkl')
self.signal[number_i] = saved_stastics.read_my_signal_results()
number_i +=1
def __init__(self, batch_size, image_size, path_size):
self.dataroot = "..\\dataset\\CostMatrix\\"
self.signalroot = "..\\dataset\\saved_stastics\\"
self.read_all_flag = 0
self.read_record = 0
self.folder_pointer = 0
self.batch_size = batch_size
self.img_size = image_size
self.path_size = path_size
self.input_image = np.zeros((batch_size, 1, image_size, image_size))
self.input_path = np.zeros((batch_size, path_size))
self.all_dir_list = os.listdir(self.dataroot)
self.folder_num = len(self.all_dir_list)
# create the buffer list
self.folder_list = [None] * self.folder_num
self.signal = [None] * self.folder_num
# create all the folder list and their data list
number_i = 0
# all_dir_list is subfolder list
#creat the image list point to the STASTICS TIS list
saved_stastics = MY_ANALYSIS()
#read all the folder list
for subfold in self.all_dir_list:
#if(number_i==0):
this_folder_list = os.listdir(os.path.join(self.dataroot, subfold))
this_folder_list2 = [
self.dataroot + subfold + "\\" + pointer
for pointer in this_folder_list
]
self.folder_list[number_i] = this_folder_list2
#change the dir firstly before read
saved_stastics.all_statics_dir = os.path.join(
self.signalroot, subfold, 'signals.pkl')
self.signal[number_i] = saved_stastics.read_my_signal_results()
number_i += 1
Display_STD_flag = False
Padd_zero_top = True
Display_signal_flag = False
Display_Matrix_flag = False
save_matlab_flag = True
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
video_sizeH = 450
video_sizeW = 900
#videoout = cv2.VideoWriter(save_display_dir+'0output.avi', -1, 20.0, (video_sizeW,video_sizeH))
if Display_signal_flag == True:
from analy import MY_ANALYSIS
#show the stastics results
saved_stastics = MY_ANALYSIS()
saved_stastics = saved_stastics.read_my_signal_results()
saved_stastics.display()
def tranfer2circ_padding(img):
H, W_ini = img.shape
padding = np.zeros((Padding_H, W_ini))
if Padd_zero_top == True:
img = np.append(padding, img, axis=0)
circular = tranfer_frome_rec2cir(img)
return circular
def diplay_sequence():
#show the image results
class DATA_Generator(object):
def __init__(self):
self.original_root = "../../saved_original_for_generator/"
self.data_pair1_root = "../../saved_pair1/"
self.data_pair2_root = "../../saved_pair2/"
self.data_mat_root = "../../saved_matrix/"
self.data_mat_root_origin = "../../saved_matrix_unprocessed/"
self.data_signal_root = "../../saved_stastics_for_generator/"
# noise type
self.noise_selector = [
'gauss_noise', 'gauss_noise', 'gauss_noise', 'gauss_noise'
]
self.save_matlab_root = "../../saved_matlab/"
self.self_check_path_create(self.save_matlab_root)
self.H = 1024
self.W = 832
self.matlab = Save_Signal_matlab()
# read the signals just use the existing path
self.saved_stastics = MY_ANALYSIS()
self.saved_stastics.all_statics_dir = os.path.join(
self.data_signal_root, 'signals.pkl')
self.shift_predictor = Shift_Predict()
self.path_DS = self.saved_stastics.read_my_signal_results()
self.path_DS.all_statics_dir = self.saved_stastics.all_statics_dir
if visdom_show_flag == True:
self.vis_ploter = VisdomLinePlotter()
# if there is no path, generate thsi path
def self_check_path_create(self, directory):
try:
os.stat(directory)
except:
os.mkdir(directory)
# read the original signal and then close the startign with the end
def close_the_origin_signal(self):
read_id = 0 # read pointer initialization
while (1):
OriginalpathDirlist = os.listdir(self.original_root) #
#
#read the path and Image number from the signal file
#get the Id of image which should be poibnt to
#get the path
path = self.path_DS.path_saving[read_id, :]
path_l = len(path)
long_path = np.append(path, path, axis=0)
long_path = np.append(long_path, path, axis=0)
#long_path = np.append(path[::-1],path,axis=0)
#long_path = np.append(long_path,path[::-1],axis=0)
long_path = gaussian_filter1d(long_path,
5) # the fileter parameters is 5
path_p = long_path[path_l:2 * path_l]
#change the signal too
self.path_DS.path_saving[read_id, :] = path_p
self.path_DS.save()
## validation
#steam[Len_steam-1,:,:] = original_IMG # un-correct
#steam[Len_steam-2,:,:] = Shifted_IMG # correct
#Costmatrix,shift_used = COSTMtrix.matrix_cal_corre_full_version3_2GPU(original_IMG,Shifted_IMG,0)
##Costmatrix = myfilter.gauss_filter_s (Costmatrix) # smooth matrix
#show1 = Costmatrix
#for i in range ( len(path)):
# show1[int(path[i]),i]=254
#cv2.imwrite(self.data_mat_root + str(Image_ID) +".jpg", show1)
print("[%s] is processed. test point time is [%f] " %
(read_id, 0.1))
read_id += 1
pass
def add_lines_to_matrix(self, matrix):
value = 128
H, W = matrix.shape
line_positions = np.arange(0, W - 2 * H, H)
for lines in line_positions:
for i in np.arange(0, H):
matrix[i, lines + i] = value
matrix[i, lines + i + 1] = value
matrix[i, lines + i + 3] = value
return matrix
def random_min_clip_by_row(self, min1, min2, mat):
rand = np.random.random_sample()
rand = rand * (min2 - min1) + min1
H, W = mat.shape
for i in np.arange(W):
rand = np.random.random_sample()
rand = rand * (min2 - min1) + min1
mat[:, i] = np.clip(mat[:, i], rand, 254)
return mat
def noisy(self, noise_typ, image):
if (noise_typ == "none"):
return image
if noise_typ == "gauss_noise":
row, col = image.shape
mean = 0
var = 15
sigma = var**0.5
gauss = np.random.normal(mean, sigma, (row, col))
gauss = gauss.reshape(row, col)
noisy = image + gauss
return np.clip(noisy, 0, 254)
elif noise_typ == 's&p':
row, col = image.shape
s_vs_p = 0.5
amount = 0.004
out = np.copy(image)
# Salt mode
num_salt = np.ceil(amount * image.size * s_vs_p)
coords = [
np.random.randint(0, i - 1, int(num_salt)) for i in image.shape
]
out[coords] = 1
# Pepper mode
num_pepper = np.ceil(amount * image.size * (1. - s_vs_p))
coords = [
np.random.randint(0, i - 1, int(num_pepper))
for i in image.shape
]
out[coords] = 0
return np.clip(out, 0, 254)
elif noise_typ == 'poisson':
vals = len(np.unique(image))
vals = 2**np.ceil(np.log2(vals))
noisy = np.random.poisson(image * vals) / float(vals)
return np.clip(noisy, 0, 254)
elif noise_typ == 'speckle':
row, col = image.shape
gauss = np.random.randn(row, col)
gauss = gauss.reshape(row, col)
noisy = image + image * gauss
return np.clip(noisy, 0, 254)
#the validation functionfor check the matrix and can also be used for validate the correction result
def validation_shift(self, original_IMG, Shifted_IMG, path, Image_ID):
Costmatrix, shift_used = COSTMtrix.matrix_cal_corre_full_version3_2GPU(
original_IMG, Shifted_IMG, 0)
if Clip_matrix_flag == True:
#Costmatrix = np.clip(Costmatrix, 20,254)
Costmatrix = self.random_min_clip_by_row(5, 30, Costmatrix)
Shifted_IMG2 = Shifted_IMG
shift = self.shift_predictor.predict(original_IMG, Shifted_IMG,
Shifted_IMG2)
path_deep = shift + path * 0
##middle_point = PATH.calculate_ave_mid(mat)
#path1,path_cost1=PATH.search_a_path(mat,start_point) # get the path and average cost of the path
show1 = np.zeros((Costmatrix.shape[0], Costmatrix.shape[1], 3))
cv2.imwrite(self.data_mat_root_origin + str(Image_ID) + ".jpg", show1)
show1[:, :, 0] = Costmatrix
show1[:, :, 1] = Costmatrix
show1[:, :, 2] = Costmatrix
for i in range(len(path)):
painter = min(path[i], Window_LEN - 1)
#painter2= min(path_tradition[i],Window_LEN-1)
painter3 = min(path_deep[i], Window_LEN - 1)
show1[int(painter), i, :] = [255, 255, 255]
#show1[int(painter2),i,:]=[254,0,0]
show1[int(painter3), i, :] = [0, 0, 254]
# save the matrix to fil dir
cv2.imwrite(self.data_mat_root + str(Image_ID) + ".jpg", show1)
def validation(self, original_IMG, Shifted_IMG, path, Image_ID):
#Costmatrix,shift_used = COSTMtrix.matrix_cal_corre_full_version3_2GPU(original_IMG,Shifted_IMG,0)
#Costmatrix,shift_used = COSTMtrix.matrix_cal_corre_full_version3_2GPU(original_IMG,Shifted_IMG,0)
Costmatrix, shift_used = COSTMtrix.matrix_cal_corre_full_version3_2GPU(
original_IMG, Shifted_IMG, 0)
#Costmatrix=cv2.blur(Costmatrix,(5,5))
Costmatrix = myfilter.gauss_filter_s(Costmatrix) # smooth matrix
# Costmatrix =cv2.GaussianBlur(Costmatrix,(5,5),0)
# down sample the materix and up sample
#Hm,Wm= Costmatrix.shape
#Costmatrix = cv2.resize(Costmatrix, (int(Wm/2),int(Hm/2)), interpolation=cv2.INTER_LINEAR)
#Costmatrix = cv2.resize(Costmatrix, (Wm,Hm), interpolation=cv2.INTER_LINEAR)
if Clip_matrix_flag == True:
Costmatrix = np.clip(Costmatrix, 20, 254)
#Costmatrix=self.random_min_clip_by_row(5,30,Costmatrix)
#Costmatrix = self.add_lines_to_matrix(Costmatrix)
#Costmatrix=np.clip(Costmatrix, 20, 255)
# Costmatrix = myfilter.gauss_filter_s(Costmatrix) # smooth matrix
#tradition way to find path
##middle_point = PATH.calculate_ave_mid(mat)
#path1,path_cost1=PATH.search_a_path(mat,start_point) # get the path and average cost of the path
show1 = np.zeros((Costmatrix.shape[0], Costmatrix.shape[1], 3))
show1[:, :, 0] = Costmatrix
show1[:, :, 1] = Costmatrix
show1[:, :, 2] = Costmatrix
for i in range(len(path)):
painter = np.clip(path[i], 1, Window_LEN - 2)
show1[int(painter), i, :] = show1[int(painter) - 1,
i, :] = [254, 254, 254]
if Show_nurd_compare == True:
start_point = PATH.find_the_starting(
Costmatrix) # starting point for path searching
path_tradition, pathcost1 = PATH.search_a_path(
Costmatrix,
start_point) # get the path and average cost of the path
#path_tradition=(path_tradition -Window_LEN/2)* Down_sample_F2 +Window_LEN/2
#path_deep,path_cost2=PATH.search_a_path_Deep_Mat2longpath(Costmatrix) # get the path and average cost of the path
path_deep, path_cost2 = PATH.search_a_path_GPU(
Costmatrix) # get the path and average cost of the path
#path_deep=(path_deep -Window_LEN/2)* Down_sample_F2 +Window_LEN/2
path_deep = gaussian_filter1d(path_deep, 3) # smooth the path
show1 = np.clip(show1, 1, 190)
for i in range(len(path)):
painter = np.clip(path[i], 1, Window_LEN - 2)
show1[int(painter), i, :] = show1[int(painter) - 1,
i, :] = [254, 254, 254]
painter2 = np.clip(path_tradition[i], 1, Window_LEN - 2)
painter3 = np.clip(path_deep[i], 1, Window_LEN - 2)
show1[int(painter2), i, :] = show1[int(painter2) - 1,
i, :] = [0, 254, 0]
show1[int(painter3), i, :] = show1[int(painter3) - 1,
i, :] = [0, 0, 254]
# save the matrix to fil dir
cv2.imwrite(self.data_mat_root_origin + str(Image_ID) + ".jpg",
Costmatrix)
cv2.imwrite(self.data_mat_root + str(Image_ID) + ".jpg", show1)
# show the signal comparison in visdom
if visdom_show_flag == True:
x = np.arange(0, len(path))
self.vis_ploter.plot_multi_arrays_append(x,
path,
title_name=str(Image_ID),
legend='truth')
self.vis_ploter.plot_multi_arrays_append(x,
path_deep,
title_name=str(Image_ID),
legend='Deep Learning')
self.vis_ploter.plot_multi_arrays_append(x,
path_tradition,
title_name=str(Image_ID),
legend='Traditional')
# save comparison signals to matlab
if Save_matlab_flag == True:
self.matlab.buffer_4(Image_ID, path, path_deep, path_tradition)
self.matlab.save_mat()
pass
def generate_NURD(self):
#read one from the original
#random select one IMG frome the oringinal
read_id = 0
Len_steam = 5
steam = np.zeros((Len_steam, self.H, self.W)) # create video buffer
num_path, path_len = self.path_DS.path_saving.shape
for read_id in range(num_path):
OriginalpathDirlist = os.listdir(self.original_root) #
sample = random.sample(OriginalpathDirlist, 1) #
Sample_path = self.original_root + sample[0]
original_IMG = cv2.imread(Sample_path)
path = self.path_DS.path_saving[read_id, :]
self.W = len(path)
original_IMG = cv2.cvtColor(original_IMG, cv2.COLOR_BGR2GRAY)
original_IMG = cv2.resize(original_IMG, (self.W, self.H),
interpolation=cv2.INTER_AREA)
#read the path and Image number from the signal file
#get the Id of image which should be poibnt to
Image_ID = int(
self.path_DS.signals[Save_signal_enum.image_iD.value, read_id])
#get the path
#path = signal.resample(path, self.W)#resample the path
if NURD_remove_shift_flag == True:
path = path - (np.mean(path) - Window_LEN / 2)
#path= path*0+ int(Window_LEN/2 )
Dice = int(np.random.random_sample() * 100)
if Dice % 2 == 0 or Use_random_NURD == True:
path = path * 0 + int(Window_LEN / 2)
fact1 = int(np.random.random_sample() * 20) + 20
fact2 = np.random.random_sample()
random_NURD = np.random.random_sample(
fact1) * 20 - 10 + fact2 * 20 - 10
random_NURD = signal.resample(random_NURD,
self.W) #resample the path
#random_NURD = np.random.random_sample(self.W)*30-10 + np.random.random_sample()*40-20
random_NURD = gaussian_filter1d(random_NURD,
5) # smooth the path
path = path + random_NURD
Low_path = int(35 / Down_sample_F2)
#Low_path =Low_path.astype(int)
Downsample_bias = 35 - Low_path * Down_sample_F2
path = np.clip(path, 0, Window_LEN - 1)
# create the shifted image
Shifted_IMG = VIDEO_PEOCESS.de_distortion(original_IMG, path,
Image_ID, 0)
#path = path -0.5* Downsample_bias
path = gaussian_filter1d(path, 3) # smooth the path
path = np.clip(path, 0, Window_LEN - 1)
self.path_DS.path_saving[read_id, :] = path
self.path_DS.save()
if add_noise_flag == True:
#original_IMG = Basic_Operator.add_speckle_or_not(original_IMG)
#Shifted_IMG = Basic_Operator.add_speckle_or_not(Shifted_IMG)
noise_it = np.random.random_sample() * 100
noise_type = str(self.noise_selector[int(noise_it) % 4])
#noise_type = "gauss_noise"
original_IMG = self.noisy(noise_type, original_IMG)
noise_it = np.random.random_sample() * 100
noise_type = str(self.noise_selector[int(noise_it) % 4])
Shifted_IMG = self.noisy(noise_type, Shifted_IMG)
#Shifted_IMG = self.noisy(noise_type,Shifted_IMG)
# save all the result
cv2.imwrite(self.data_pair1_root + str(Image_ID) + ".jpg",
original_IMG)
cv2.imwrite(self.data_pair2_root + str(Image_ID) + ".jpg",
Shifted_IMG)
## validation
#self.validation(original_IMG,Shifted_IMG,path,Image_ID)
#steam[Len_steam-1,:,:] = original_IMG # un-correct
#steam[Len_steam-2,:,:] = Shifted_IMG # correct
#Costmatrix,shift_used = COSTMtrix.matrix_cal_corre_full_version3_2GPU(original_IMG,Shifted_IMG,0)
#Costmatrix = myfilter.gauss_filter_s (Costmatrix) # smooth matrix
#show1 = Costmatrix
#for i in range ( len(path)):
# show1[int(path[i]),i]=254
#cv2.imwrite(self.data_mat_root + str(Image_ID) +".jpg", show1)
print("[%s] is processed. test point time is [%f] " %
(read_id, 0.1))
#read_id +=1
def generate_overall_shifting(self):
#read one from the original
#random select one IMG frome the oringinal
read_id = 0
Len_steam = 5
#steam=np.zeros((Len_steam,self.H,self.W)) # create video buffer
while (1):
random_shifting = np.random.random_sample(
) * Overall_shiftting_WinLen
#random_shifting = random.random() * Overall_shiftting_WinLen
OriginalpathDirlist = os.listdir(self.original_root) #
sample = random.sample(OriginalpathDirlist, 1) #
Sample_path = self.original_root + sample[0]
original_IMG = cv2.imread(Sample_path)
original_IMG = cv2.cvtColor(original_IMG, cv2.COLOR_BGR2GRAY)
original_IMG = cv2.resize(original_IMG, (self.W, self.H),
interpolation=cv2.INTER_AREA)
#original_IMG = cv2.resize(original_IMG, (self.W,self.H), interpolation=cv2.INTER_AREA)
H, W = original_IMG.shape
#read the path and Image number from the signal file
#get the Id of image which should be poibnt to
Image_ID = int(
self.path_DS.signals[Save_signal_enum.image_iD.value, read_id])
#get the path
path = self.path_DS.path_saving[read_id, :]
#change the signal too
self.path_DS.path_saving[read_id, :] = path * 0 + random_shifting
path = signal.resample(path,
W) * 0 + random_shifting #resample the path
#resave the signal
# create the shifted image
Shifted_IMG = VIDEO_PEOCESS.de_distortion(original_IMG, path,
Image_ID, 0)
# save all the result
cv2.imwrite(self.data_pair1_root + str(Image_ID) + ".jpg",
original_IMG)
cv2.imwrite(self.data_pair2_root + str(Image_ID) + ".jpg",
Shifted_IMG)
self.path_DS.save()
#self.validation_shift(original_IMG,Shifted_IMG,path,Image_ID)
## validation
#steam[Len_steam-1,:,:] = original_IMG # un-correct
#steam[Len_steam-2,:,:] = Shifted_IMG # correct
#Costmatrix,shift_used = COSTMtrix.matrix_cal_corre_full_version3_2GPU(original_IMG,Shifted_IMG,0)
##Costmatrix = myfilter.gauss_filter_s (Costmatrix) # smooth matrix
#show1 = Costmatrix
#for i in range ( len(path)):
# show1[int(path[i]),i]=254
#cv2.imwrite(self.data_mat_root + str(Image_ID) +".jpg", show1)
print("[%s] is processed. test point time is [%f] " %
(read_id, 0.1))
read_id += 1
# generate the OCT iamge with combination of NURD and group shifting
def generate_NURD_overall_shifting(self):
#read one from the original
#random select one IMG frome the oringinal
read_id = 0 # read pointer initialization
Len_steam = 5 # create the buffer for validation
steam = np.zeros((Len_steam, self.H, self.W)) # create video buffer
growing = 0
while (1):
#list all the picture for video generating, ensure the original folder has only one image
OriginalpathDirlist = os.listdir(self.original_root)
sample = random.sample(OriginalpathDirlist,
1) # ramdom choose the name in folder list
Sample_path = self.original_root + sample[
0] # create the reading path this radom picture
original_IMG = cv2.imread(Sample_path) # get this image
original_IMG = cv2.cvtColor(original_IMG,
cv2.COLOR_BGR2GRAY) # to gray
original_IMG = cv2.resize(original_IMG, (self.W, self.H),
interpolation=cv2.INTER_LINEAR)
#read the path and Image number from the signal file
#get the Id of image which should be poibnt to
Image_ID = int(
self.path_DS.signals[Save_signal_enum.image_iD.value, read_id])
#get the path
path = self.path_DS.path_saving[read_id, :]
path = signal.resample(path, self.W) #resample the path
if NURD_remove_shift_flag == True:
path = path - (np.mean(path) - Window_LEN / 2)
#path= path*0+ int(Window_LEN/2 )
Dice = int(np.random.random_sample() * 100)
if Dice % 2 == 0 or Use_random_NURD == True:
path = path * 0 + int(Window_LEN / 2)
fact1 = int(np.random.random_sample() * 50) + 20
fact2 = np.random.random_sample()
random_NURD = np.random.random_sample(fact1) * 30 - 15
random_NURD = signal.resample(random_NURD,
self.W) #resample the path
#random_NURD = np.random.random_sample(self.W)*30-10 + np.random.random_sample()*40-20
random_NURD = gaussian_filter1d(random_NURD,
10) # smooth the path
path = path + random_NURD
# exragene for diaplay
#path = (path -np.mean(path))*0.6+np.mean(path)
#overall_shifting = Image_ID
#overall_shifting = min(overall_shifting,self.W/2) # limit the shifting here, maybe half the lenghth is sufficient for the combination
#overall_shifting = min(overall_shifting,self.W/2) # limit the shifting here, maybe half the lenghth is sufficient for the combination
random_shifting = np.random.random_sample(
) * Overall_shiftting_WinLen / 2
#Combine the overall shifting with NURD
#path = path + overall_shifting
path = path + random_shifting
# create the shifted image
Shifted_IMG = VIDEO_PEOCESS.de_distortion(original_IMG, path,
Image_ID, 0)
#modify for training ground trueth
self.path_DS.path_saving[read_id, :] = self.path_DS.path_saving[
read_id, :] * 0 + random_shifting
self.path_DS.save()
# add noise to image pair for validation
if add_noise_flag == True:
noise_type = str(self.noise_selector[int(Image_ID) % 4])
#noise_type = "gauss_noise"
original_IMG = self.noisy(noise_type, original_IMG)
Shifted_IMG = self.noisy(noise_type, Shifted_IMG)
if Image_ID < 100:
grower = +np.random.random_sample() * np.random.random_sample(
) * 2
else:
grower = +np.random.random_sample() * np.random.random_sample(
) * 1
growing += grower
# addistional shift
Shifted_IMG = np.roll(Shifted_IMG, int(growing), axis=1)
# save all the result
cv2.imwrite(self.data_pair1_root + str(Image_ID) + ".jpg",
original_IMG)
cv2.imwrite(self.data_pair2_root + str(Image_ID) + ".jpg",
Shifted_IMG)
# save generate information: the NURD and shift used for generating img pair to matlab
if Save_matlab_flag == True:
self.matlab.buffer_overall_shift_NURD(Image_ID,
overall_shifting, path)
self.matlab.save_mat_infor_of_over_allshift_with_NURD()
self.matlab.save_pkl_infor_of_over_allshift_with_NURD()
pass
## validation
if validation_flag == True:
self.validation_shift(original_IMG, Shifted_IMG, path,
Image_ID)
#self.validation(original_IMG,Shifted_IMG,path,Image_ID)
#steam[Len_steam-1,:,:] = original_IMG # un-correct
#steam[Len_steam-2,:,:] = Shifted_IMG # correct
#Costmatrix,shift_used = COSTMtrix.matrix_cal_corre_full_version3_2GPU(original_IMG,Shifted_IMG,0)
#Costmatrix = myfilter.gauss_filter_s (Costmatrix) # smooth matrix
#show1 = Costmatrix
#for i in range ( len(path)):
# show1[int(path[i]),i]=254
#cv2.imwrite(self.data_mat_root + str(Image_ID) +".jpg", show1)
print("[%s] is processed. test point time is [%f] " %
(read_id, 0.1))
read_id += 1
class DATA_Generator(object):
def __init__(self):
self.original_root = "../../saved_original_for_generator/"
self.data_pair1_root = "../../saved_pair1/"
self.data_pair2_root = "../../saved_pair2/"
self.data_mat_root = "../../saved_matrix/"
self.data_mat_root_origin = "../../saved_matrix_unprocessed/"
self.data_signal_root = "../../saved_stastics_for_generator/"
self.H = 1024
self.W = 780
# read the signals just use the existing path
self.saved_stastics = MY_ANALYSIS()
self.saved_stastics.all_statics_dir = os.path.join(self.data_signal_root, 'signals.pkl')
self.path_DS = self.saved_stastics.read_my_signal_results()
self.path_DS.all_statics_dir = self.saved_stastics.all_statics_dir
#the validation functionfor check the matrix and can also be used for validate the correction result
def validation(self,original_IMG,Shifted_IMG,path,Image_ID):
#Costmatrix,shift_used = COSTMtrix.matrix_cal_corre_full_version3_2GPU(original_IMG,Shifted_IMG,0)
Costmatrix,shift_used = COSTMtrix.matrix_cal_corre_block_version3_3GPU(original_IMG,Shifted_IMG,0)
# Costmatrix = myfilter.gauss_filter_s(Costmatrix) # smooth matrix
#tradition way to find path
start_point= PATH.find_the_starting(Costmatrix) # starting point for path searching
#path_tradition,pathcost1 = PATH.search_a_path(Costmatrix,start_point) # get the path and average cost of the path
#path_deep,path_cost2=PATH.search_a_path_Deep_Mat2longpath(Costmatrix) # get the path and average cost of the path
path_deep,path_cost2=PATH.search_a_path_deep_multiscal_small_window(Costmatrix) # get the path and average cost of the path
path_deep = gaussian_filter1d(path_deep,3) # smooth the path
##middle_point = PATH.calculate_ave_mid(mat)
#path1,path_cost1=PATH.search_a_path(mat,start_point) # get the path and average cost of the path
show1 = Costmatrix
cv2.imwrite(self.data_mat_root_origin + str(Image_ID) +".jpg", show1)
for i in range ( len(path)):
painter = min(path[i],Window_LEN-1)
#painter2= min(path_tradition[i],Window_LEN-1)
painter3 = min(path_deep[i],Window_LEN-1)
show1[int(painter),i]=128
#show1[int(painter2),i]=128
show1[int(painter3),i]=254
cv2.imwrite( self.data_mat_root + str(Image_ID) +".jpg", show1)
def generate_NURD(self):
#read one from the original
#random select one IMG frome the oringinal
read_id = 0
Len_steam =5
steam=np.zeros((Len_steam,self.H,self.W)) # create video buffer
while (1):
OriginalpathDirlist = os.listdir(self.original_root) #
sample = random.sample(OriginalpathDirlist, 1) #
Sample_path = self.original_root + sample[0]
original_IMG = cv2.imread(Sample_path)
original_IMG = cv2.cvtColor(original_IMG, cv2.COLOR_BGR2GRAY)
original_IMG = cv2.resize(original_IMG, (self.W,self.H), interpolation=cv2.INTER_AREA)
#read the path and Image number from the signal file
#get the Id of image which should be poibnt to
Image_ID = int( self.path_DS.signals[Save_signal_enum.image_iD.value, read_id])
#get the path
path = self.path_DS.path_saving[read_id,:]
path = signal.resample(path, self.W)#resample the path
# create the shifted image
Shifted_IMG = VIDEO_PEOCESS.de_distortion(original_IMG,path,Image_ID,0)
# save all the result
cv2.imwrite(self.data_pair1_root + str(Image_ID) +".jpg", original_IMG)
cv2.imwrite(self.data_pair2_root + str(Image_ID) +".jpg", Shifted_IMG)
## validation
self.validation(original_IMG,Shifted_IMG,path,Image_ID)
#steam[Len_steam-1,:,:] = original_IMG # un-correct
#steam[Len_steam-2,:,:] = Shifted_IMG # correct
#Costmatrix,shift_used = COSTMtrix.matrix_cal_corre_full_version3_2GPU(original_IMG,Shifted_IMG,0)
#Costmatrix = myfilter.gauss_filter_s (Costmatrix) # smooth matrix
#show1 = Costmatrix
#for i in range ( len(path)):
# show1[int(path[i]),i]=254
#cv2.imwrite(self.data_mat_root + str(Image_ID) +".jpg", show1)
print ("[%s] is processed. test point time is [%f] " % (read_id ,0.1))
read_id +=1
def generate_overall_shifting(self):
#read one from the original
#random select one IMG frome the oringinal
read_id = 0
Len_steam =5
#steam=np.zeros((Len_steam,self.H,self.W)) # create video buffer
while (1):
random_shifting = random.random() * Overall_shiftting_WinLen
OriginalpathDirlist = os.listdir(self.original_root) #
sample = random.sample(OriginalpathDirlist, 1) #
Sample_path = self.original_root + sample[0]
original_IMG = cv2.imread(Sample_path)
original_IMG = cv2.cvtColor(original_IMG, cv2.COLOR_BGR2GRAY)
#original_IMG = cv2.resize(original_IMG, (self.W,self.H), interpolation=cv2.INTER_AREA)
H,W = original_IMG.shape
#read the path and Image number from the signal file
#get the Id of image which should be poibnt to
Image_ID = int( self.path_DS.signals[Save_signal_enum.image_iD.value, read_id])
#get the path
path = self.path_DS.path_saving[read_id,:]
#change the signal too
self.path_DS.path_saving[read_id,:] = path* 0 + random_shifting
path = signal.resample(path, W)*0 + random_shifting #resample the path
#resave the signal
# create the shifted image
Shifted_IMG = VIDEO_PEOCESS.de_distortion(original_IMG,path,Image_ID,0)
# save all the result
cv2.imwrite(self.data_pair1_root + str(Image_ID) +".jpg", original_IMG)
cv2.imwrite(self.data_pair2_root + str(Image_ID) +".jpg", Shifted_IMG)
self.path_DS.save()
self.validation(original_IMG,Shifted_IMG,path,Image_ID)
## validation
#steam[Len_steam-1,:,:] = original_IMG # un-correct
#steam[Len_steam-2,:,:] = Shifted_IMG # correct
#Costmatrix,shift_used = COSTMtrix.matrix_cal_corre_full_version3_2GPU(original_IMG,Shifted_IMG,0)
##Costmatrix = myfilter.gauss_filter_s (Costmatrix) # smooth matrix
#show1 = Costmatrix
#for i in range ( len(path)):
# show1[int(path[i]),i]=254
#cv2.imwrite(self.data_mat_root + str(Image_ID) +".jpg", show1)
print ("[%s] is processed. test point time is [%f] " % (read_id ,0.1))
read_id +=1
def generate_NURD_overall_shifting(self):
pass
class DATA_augmentor(object):
def __init__(self):
self.original_root = "../../saved_original_for_generator/"
self.data_pair1_root = "../../saved_pair1/"
self.data_pair2_root = "../../saved_pair2/"
self.data_mat_root = "../../saved_matrix/"
self.data_mat_root_origin = "../../saved_matrix_unprocessed/"
self.data_mat_root_augmented = "../../saved_matrix_augmented/"
self.data_signal_root = "../../saved_stastics_for_generator/"
if not os.path.exists(self.data_mat_root_augmented):
os.mkdir(self.data_mat_root_augmented)
print("Directory ", self.data_mat_root_augmented, " Created ")
else:
print("Directory ", self.data_mat_root_augmented,
" already exists")
self.H = 1024
self.W = 780
# read the signals just use the existing path
self.saved_stastics = MY_ANALYSIS()
self.saved_stastics.all_statics_dir = os.path.join(
self.data_signal_root, 'signals.pkl')
self.path_DS = self.saved_stastics.read_my_signal_results()
self.path_DS.all_statics_dir = self.saved_stastics.all_statics_dir
#the validation functionfor check the matrix and can also be used for validate the correction result
def validation(self, original_IMG, Shifted_IMG, path, Image_ID):
Costmatrix, shift_used = COSTMtrix.matrix_cal_corre_full_version3_2GPU(
original_IMG, Shifted_IMG, 0)
# Costmatrix = myfilter.gauss_filter_s(Costmatrix) # smooth matrix
#tradition way to find path
start_point = PATH.find_the_starting(
Costmatrix) # starting point for path searching
#path_tradition,pathcost1 = PATH.search_a_path(Costmatrix,start_point) # get the path and average cost of the path
path_deep, path_cost2 = PATH.search_a_path_Deep_Mat2longpath(
Costmatrix) # get the path and average cost of the path
path_deep = gaussian_filter1d(path_deep, 3) # smooth the path
##middle_point = PATH.calculate_ave_mid(mat)
#path1,path_cost1=PATH.search_a_path(mat,start_point) # get the path and average cost of the path
show1 = Costmatrix
cv2.imwrite(self.data_mat_root_origin + str(Image_ID) + ".jpg", show1)
for i in range(len(path)):
painter = min(path[i], Window_LEN - 1)
#painter2= min(path_tradition[i],Window_LEN-1)
painter3 = min(path_deep[i], Window_LEN - 1)
show1[int(painter), i] = 128
#show1[int(painter2),i]=128
show1[int(painter3), i] = 254
cv2.imwrite(self.data_mat_root + str(Image_ID) + ".jpg", show1)
def noisy(self, noise_typ, image):
if noise_typ == "gauss":
row, col = image.shape
mean = 0
var = 50
sigma = var**0.5
gauss = np.random.normal(mean, sigma, (row, col))
gauss = gauss.reshape(row, col)
noisy = image + gauss
return noisy
elif noise_typ == "s&p":
row, col = image.shape
s_vs_p = 0.5
amount = 0.004
out = np.copy(image)
# Salt mode
num_salt = np.ceil(amount * image.size * s_vs_p)
coords = [
np.random.randint(0, i - 1, int(num_salt)) for i in image.shape
]
out[coords] = 1
# Pepper mode
num_pepper = np.ceil(amount * image.size * (1. - s_vs_p))
coords = [
np.random.randint(0, i - 1, int(num_pepper))
for i in image.shape
]
out[coords] = 0
return out
elif noise_typ == "poisson":
vals = len(np.unique(image))
vals = 2**np.ceil(np.log2(vals))
noisy = np.random.poisson(image * vals) / float(vals)
return noisy
elif noise_typ == "speckle":
row, col = image.shape
gauss = np.random.randn(row, col)
gauss = gauss.reshape(row, col)
noisy = image + image * gauss
return noisy
#def add_gaussian_noise(self,X_imgs):
# gaussian_noise_imgs = []
# row, col, _ = X_imgs[0].shape
# # Gaussian distribution parameters
# mean = 0
# var = 0.1
# sigma = var ** 0.5
# for X_img in X_imgs:
# gaussian = np.random.random((row, col, 1)).astype(np.float32)
# gaussian = np.concatenate((gaussian, gaussian, gaussian), axis = 2)
# gaussian_img = cv2.addWeighted(X_img, 0.75, 0.25 * gaussian, 0.25, 0)
# gaussian_noise_imgs.append(gaussian_img)
# gaussian_noise_imgs = np.array(gaussian_noise_imgs, dtype = np.float32)
# return gaussian_noise_imgs
def augment_gauss_noise(self):
#read one from the original
noise_selector = ["gauss", "s&p", "poisson", "speckle"]
for img in os.listdir(self.data_mat_root_origin):
a, b = os.path.splitext(img)
if b == ".jpg":
original_IMG = cv2.imread(self.data_mat_root_origin + img)
original_IMG = cv2.cvtColor(original_IMG, cv2.COLOR_BGR2GRAY)
Gauss_IMG = self.noisy(noise_selector[int(a) % 4],
original_IMG)
# save all the result
#cv2.imwrite(self.data_pair1_root + str(Image_ID) +".jpg", original_IMG)
cv2.imwrite(self.data_mat_root_augmented + a + ".jpg",
Gauss_IMG)
print("[%s] is processed. test point time is [%f] " % (a, 0.1))
def augment_blur(self):
#read one from the original
noise_selector = ["gauss", "s&p", "poisson", "speckle"]
for img in os.listdir(self.data_mat_root_origin):
a, b = os.path.splitext(img)
if b == ".jpg":
original_IMG = cv2.imread(self.data_mat_root_origin + img)
original_IMG = cv2.cvtColor(original_IMG, cv2.COLOR_BGR2GRAY)
Blur_IMG = cv2.blur(original_IMG, (5, 5))
# save all the result
#cv2.imwrite(self.data_pair1_root + str(Image_ID) +".jpg", original_IMG)
cv2.imwrite(self.data_mat_root_augmented + a + ".jpg",
Blur_IMG)
print("[%s] is processed. test point time is [%f] " % (a, 0.1))
def add_lines_to_matrix(self, matrix):
value = 128
H, W = matrix.shape
line_positions = np.arange(0, W - 2 * H, H)
for lines in line_positions:
for i in np.arange(0, H):
matrix[i, lines + i] = value
return matrix
def augment_add_lines(self):
#read one from the original
for img in os.listdir(self.data_mat_root_origin):
a, b = os.path.splitext(img)
if b == ".jpg":
original_IMG = cv2.imread(self.data_mat_root_origin + img)
original_IMG = cv2.cvtColor(original_IMG, cv2.COLOR_BGR2GRAY)
Add_line_IMG = self.add_lines_to_matrix(original_IMG)
# save all the result
#cv2.imwrite(self.data_pair1_root + str(Image_ID) +".jpg", original_IMG)
cv2.imwrite(self.data_mat_root_augmented + a + ".jpg",
Add_line_IMG)
print("[%s] is processed. test point time is [%f] " % (a, 0.1))
def augment_clip(self):
#read one from the original
for img in os.listdir(self.data_mat_root_origin):
a, b = os.path.splitext(img)
if b == ".jpg":
original_IMG = cv2.imread(self.data_mat_root_origin + img)
original_IMG = cv2.cvtColor(original_IMG, cv2.COLOR_BGR2GRAY)
clip_line_IMG = np.clip(original_IMG, 20, 255)
# save all the result
#cv2.imwrite(self.data_pair1_root + str(Image_ID) +".jpg", original_IMG)
cv2.imwrite(self.data_mat_root_augmented + a + ".jpg",
clip_line_IMG)
print("[%s] is processed. test point time is [%f] " % (a, 0.1))