def get_batch(self, batch, train=True):
filenames = []
images = []
edgemaps = []
double_edge = []
edgemaps_4 = []
edgemaps_8 = []
edgemaps_16 = []
chanel1 = []
chanel2 = []
chanel3 = []
chanel4 = []
chanel5 = []
chanel6 = []
chanel7 = []
chanel8 = []
chanelfuse = []
chanel = [[] for i in range(8)]
for idx, b in enumerate(batch):
if train:
index = self.X_train.index(b)
im = Image.open(self.X_train[index])
dou_path = os.path.join(self.double_edge_file, self.Y_train[index].split('/')[-1])
dou_em = Image.open(dou_path)
# 在这里获取8张图，从左上角按照顺时针顺序,返回的是一个长度为8的列表
relation_8_map = gen_8_map(dou_em)
for i in range(8):
relation_8_map[i] = Image.fromarray(relation_8_map[i].astype('uint8')).convert('RGB')
weight = im.size[0]
height = im.size[1]
if im.size[0] > 320 and im.size[1] > 320:
w_centry = weight // 2
h_centry = height // 2
if w_centry > 160:
range_w = random.randint(0, w_centry - 160)
else:
range_w = 0
if h_centry > 160:
range_h = random.randint(0, h_centry - 160)
else:
range_h = 0
# 决定图像是否加压缩
if random.randint(0, 20) == 1:
try:
mb = random.randint(30, 100)
path = compress_image(infile=self.X_train[index], mb=mb)
im = Image.open(path)
except:
traceback.print_exc()
# 决定图像是否加模糊
if random.randint(0, 20) == 1:
m = random.randint(0, 5)
if m == 0:
r = random.randint(1, 3)
im = im.filter(ImageFilter.GaussianBlur(radius=r))
else:
pass
if random.randint(0, 1) == 0:
im = im.crop((w_centry - 160 + range_w, h_centry - 160 + range_h, w_centry + 160 + range_w,
h_centry + 160 + range_h))
dou_em = dou_em.crop(
(w_centry - 160 + range_w, h_centry - 160 + range_h, w_centry + 160 + range_w,
h_centry + 160 + range_h))
for i in range(8):
relation_8_map[i] = relation_8_map[i].crop(
(w_centry - 160 + range_w, h_centry - 160 + range_h, w_centry + 160 + range_w,
h_centry + 160 + range_h))
else:
im = im.crop((w_centry - 160 - range_w, h_centry - 160 - range_h, w_centry + 160 - range_w,
h_centry + 160 - range_h))
dou_em = dou_em.crop(
(w_centry - 160 - range_w, h_centry - 160 - range_h, w_centry + 160 - range_w,
h_centry + 160 - range_h))
for i in range(8):
relation_8_map[i] = relation_8_map[i].crop(
(w_centry - 160 - range_w, h_centry - 160 - range_h, w_centry + 160 - range_w,
h_centry + 160 - range_h))
# 决定是否flip 旋转
random_aug = random.randint(0, 20)
try:
if random_aug < 5:
if random_aug == 0:
# 旋转90
im = im.transpose(Image.ROTATE_90)
dou_em = dou_em.transpose(Image.ROTATE_90)
for i in range(8):
relation_8_map[i] = relation_8_map[i].transpose(Image.ROTATE_90)
elif random_aug == 1:
# 旋转180
im = im.transpose(Image.ROTATE_180)
dou_em = dou_em.transpose(Image.ROTATE_180)
for i in range(8):
relation_8_map[i] = relation_8_map[i].transpose(Image.ROTATE_180)
elif random_aug == 2:
# 旋转270
im = im.transpose(Image.ROTATE_270)
dou_em = dou_em.transpose(Image.ROTATE_270)
for i in range(8):
relation_8_map[i] = relation_8_map[i].transpose(Image.ROTATE_270)
elif random_aug == 3:
# 左右呼唤
im = im.transpose(Image.FLIP_LEFT_RIGHT)
dou_em = dou_em.transpose(Image.FLIP_LEFT_RIGHT)
for i in range(8):
relation_8_map[i] = relation_8_map[i].transpose(Image.FLIP_LEFT_RIGHT)
elif random_aug == 4:
# 左右呼唤
im = im.transpose(Image.FLIP_TOP_BOTTOM)
dou_em = dou_em.transpose(Image.FLIP_TOP_BOTTOM)
for i in range(8):
relation_8_map[i] = relation_8_map[i].transpose(Image.FLIP_TOP_BOTTOM)
except:
traceback.print_exc()
else:
# 决定图像是否加压缩
if random.randint(0, 20) == 1:
# o_size = get_size(self.X_train[index])
try:
mb = random.randint(30, 100)
path = compress_image(infile=self.X_train[index], mb=mb)
im = Image.open(path)
except:
traceback.print_exc()
# 决定图像是否加模糊
if random.randint(0, 20) == 1:
m = random.randint(0, 5)
if m == 0:
r = random.randint(1, 3)
im = im.filter(ImageFilter.GaussianBlur(radius=r))
else:
pass
# x = random.randint(0, 200)
# y = random.randint(0, 200)
# bounds_x = random.randint(20, 100)
# bounds_y = random.randint(20, 100)
# bounds = (x, y, x + bounds_x, y + bounds_y)
# r = random.randint(1, 3)
# im = im.filter(MyGaussianBlur(radius=r, bounds=bounds))
im = im.crop((0, 0, self.image_height, self.image_width))
dou_em = dou_em.crop((0, 0, self.image_height, self.image_width))
for i in range(8):
relation_8_map[i] = relation_8_map[i].crop((0, 0, self.image_height, self.image_width))
# chanel_1_1 = chanel_1_1.crop((0, 0, self.image_height, self.image_width))
# chanel_10 = chanel_10.crop((0, 0, self.image_height, self.image_width))
# chanel_11 = chanel_11.crop((0, 0, self.image_height, self.image_width))
# chanel0_1 = chanel0_1.crop((0, 0, self.image_height, self.image_width))
# chanel01 = chanel01.crop((0, 0, self.image_height, self.image_width))
# chanel1_1 = chanel1_1.crop((0, 0, self.image_height, self.image_width))
# chanel10 = chanel10.crop((0, 0, self.image_height, self.image_width))
# chanel11 = chanel11.crop((0, 0, self.image_height, self.image_width))
# print(im)
im = np.array(im, dtype=np.float32)
im = im[..., ::-1] # RGB 2 BGR
# R = im[..., 0].mean()
# G = im[..., 1].mean()
# B = im[..., 2].mean()
# im[..., 0] -= R
# im[..., 1] -= G
# im[..., 2] -= B
# R=118.98194217348079 G=127.4061956623793 B=138.00865419127499
im[..., 0] -= 138.008
im[..., 1] -= 127.406
im[..., 2] -= 118.982
dou_chanel = [0 for i in range(8)]
for i in range(8):
dou_chanel[i] = np.array(relation_8_map[i], dtype=np.float32)
dou_chanel[i] = np.array(dou_chanel[i][:, :, 1:])
# dou_chanel[i] = dou_chanel[i] / 255
chanel[i].append(dou_chanel[i])
c_1 = dou_chanel[0][:, :, 1:]
c_2 = dou_chanel[1][:, :, 1:]
c_3 = dou_chanel[2][:, :, 1:]
c_4 = dou_chanel[3][:, :, 1:]
c_5 = dou_chanel[4][:, :, 1:]
c_6 = dou_chanel[5][:, :, 1:]
c_7 = dou_chanel[6][:, :, 1:]
c_8 = dou_chanel[7][:, :, 1:]
final_c = np.concatenate((c_1, c_2, c_3, c_4, c_5, c_6, c_7, c_8), axis=2)
chanelfuse.append(final_c)
#
# dou_em = np.array(dou_em, dtype=np.float32)
# dou_em = np.array(dou_em[:, :, :])
# dou_em = np.expand_dims(dou_em, 2)
# double_edge.append(dou_em)
dou_em = np.array(dou_em, dtype=np.float32)
# 转化为无类别的GT 100 255 为边缘
dou_em = np.where(dou_em == 50,0,dou_em)
dou_em = np.where(dou_em ==100,1,dou_em)
dou_em = np.where(dou_em == 255,1,dou_em)
dou_em = np.array(dou_em[:, :])
dou_em = np.expand_dims(dou_em, 2)
double_edge.append(dou_em)
images.append(im)
# edgemaps.append(bin_em)
# edgemaps_4.append(bin_em4)
# edgemaps_8.append(bin_em8)
# edgemaps_16.append(bin_em16)
filenames.append(self.X_train[index])
else:
index = self.X_test.index(b)
im = Image.open(self.X_test[index])
#
# index = self.X_train.index(b)
# im = Image.open(self.X_train[index])
dou_path = os.path.join(self.double_edge_file, self.Y_test[index].split('/')[-1])
dou_em = Image.open(dou_path)
# 在这里获取8张图，从左上角按照顺时针顺序,返回的是一个长度为8的列表
relation_8_map = gen_8_map(dou_em)
for i in range(8):
relation_8_map[i] = Image.fromarray(relation_8_map[i].astype('uint8')).convert('RGB')
weight = im.size[0]
height = im.size[1]
if im.size[0] > 320 and im.size[1] > 320:
w_centry = weight // 2
h_centry = height // 2
if w_centry > 160:
range_w = random.randint(0, w_centry - 160)
else:
range_w = 0
if h_centry > 160:
range_h = random.randint(0, h_centry - 160)
else:
range_h = 0
# 决定图像是否加压缩
if random.randint(0, 20) == 1:
try:
mb = random.randint(30, 100)
path = compress_image(infile=self.X_test[index], mb=mb)
im = Image.open(path)
except:
traceback.print_exc()
# 决定图像是否加模糊
if random.randint(0, 20) == 1:
m = random.randint(0, 5)
if m == 0:
r = random.randint(1, 3)
im = im.filter(ImageFilter.GaussianBlur(radius=r))
else:
pass
if random.randint(0, 1) == 0:
im = im.crop((w_centry - 160 + range_w, h_centry - 160 + range_h, w_centry + 160 + range_w,
h_centry + 160 + range_h))
dou_em = dou_em.crop(
(w_centry - 160 + range_w, h_centry - 160 + range_h, w_centry + 160 + range_w,
h_centry + 160 + range_h))
for i in range(8):
relation_8_map[i] = relation_8_map[i].crop(
(w_centry - 160 + range_w, h_centry - 160 + range_h, w_centry + 160 + range_w,
h_centry + 160 + range_h))
else:
im = im.crop((w_centry - 160 - range_w, h_centry - 160 - range_h, w_centry + 160 - range_w,
h_centry + 160 - range_h))
dou_em = dou_em.crop(
(w_centry - 160 - range_w, h_centry - 160 - range_h, w_centry + 160 - range_w,
h_centry + 160 - range_h))
for i in range(8):
relation_8_map[i] = relation_8_map[i].crop(
(w_centry - 160 - range_w, h_centry - 160 - range_h, w_centry + 160 - range_w,
h_centry + 160 - range_h))
# 决定是否flip 旋转
random_aug = random.randint(0, 20)
try:
if random_aug < 5:
if random_aug == 0:
# 旋转90
im = im.transpose(Image.ROTATE_90)
dou_em = dou_em.transpose(Image.ROTATE_90)
for i in range(8):
relation_8_map[i] = relation_8_map[i].transpose(Image.ROTATE_90)
elif random_aug == 1:
# 旋转180
im = im.transpose(Image.ROTATE_180)
dou_em = dou_em.transpose(Image.ROTATE_180)
for i in range(8):
relation_8_map[i] = relation_8_map[i].transpose(Image.ROTATE_180)
elif random_aug == 2:
# 旋转270
im = im.transpose(Image.ROTATE_270)
dou_em = dou_em.transpose(Image.ROTATE_270)
for i in range(8):
relation_8_map[i] = relation_8_map[i].transpose(Image.ROTATE_270)
elif random_aug == 3:
# 左右呼唤
im = im.transpose(Image.FLIP_LEFT_RIGHT)
dou_em = dou_em.transpose(Image.FLIP_LEFT_RIGHT)
for i in range(8):
relation_8_map[i] = relation_8_map[i].transpose(Image.FLIP_LEFT_RIGHT)
elif random_aug == 4:
# 左右呼唤
im = im.transpose(Image.FLIP_TOP_BOTTOM)
dou_em = dou_em.transpose(Image.FLIP_TOP_BOTTOM)
for i in range(8):
relation_8_map[i] = relation_8_map[i].transpose(Image.FLIP_TOP_BOTTOM)
except:
traceback.print_exc()
else:
# 决定图像是否加压缩
if random.randint(0, 20) == 1:
# o_size = get_size(self.X_train[index])
try:
mb = random.randint(30, 100)
path = compress_image(infile=self.X_test[index], mb=mb)
im = Image.open(path)
except:
traceback.print_exc()
# 决定图像是否加模糊
if random.randint(0, 20) == 1:
m = random.randint(0, 5)
if m == 0:
r = random.randint(1, 3)
im = im.filter(ImageFilter.GaussianBlur(radius=r))
else:
pass
# x = random.randint(0, 200)
# y = random.randint(0, 200)
# bounds_x = random.randint(20, 100)
# bounds_y = random.randint(20, 100)
# bounds = (x, y, x + bounds_x, y + bounds_y)
# r = random.randint(1, 3)
# im = im.filter(MyGaussianBlur(radius=r, bounds=bounds))
im = im.crop((0, 0, self.image_height, self.image_width))
dou_em = dou_em.crop((0, 0, self.image_height, self.image_width))
for i in range(8):
relation_8_map[i] = relation_8_map[i].crop((0, 0, self.image_height, self.image_width))
# chanel_1_1 = chanel_1_1.crop((0, 0, self.image_height, self.image_width))
# chanel_10 = chanel_10.crop((0, 0, self.image_height, self.image_width))
# chanel_11 = chanel_11.crop((0, 0, self.image_height, self.image_width))
# chanel0_1 = chanel0_1.crop((0, 0, self.image_height, self.image_width))
# chanel01 = chanel01.crop((0, 0, self.image_height, self.image_width))
# chanel1_1 = chanel1_1.crop((0, 0, self.image_height, self.image_width))
# chanel10 = chanel10.crop((0, 0, self.image_height, self.image_width))
# chanel11 = chanel11.crop((0, 0, self.image_height, self.image_width))
# print(im)
im = np.array(im, dtype=np.float32)
im = im[..., ::-1] # RGB 2 BGR
# R = im[..., 0].mean()
# G = im[..., 1].mean()
# B = im[..., 2].mean()
# im[..., 0] -= R
# im[..., 1] -= G
# im[..., 2] -= B
# R=118.98194217348079 G=127.4061956623793 B=138.00865419127499
im[..., 0] -= 138.008
im[..., 1] -= 127.406
im[..., 2] -= 118.982
dou_chanel = [0 for i in range(8)]
for i in range(8):
dou_chanel[i] = np.array(relation_8_map[i], dtype=np.float32)
dou_chanel[i] = np.array(dou_chanel[i][:, :, 1:])
chanel[i].append(dou_chanel[i])
c_1 = dou_chanel[0][:, :, 1:]
c_2 = dou_chanel[1][:, :, 1:]
c_3 = dou_chanel[2][:, :, 1:]
c_4 = dou_chanel[3][:, :, 1:]
c_5 = dou_chanel[4][:, :, 1:]
c_6 = dou_chanel[5][:, :, 1:]
c_7 = dou_chanel[6][:, :, 1:]
c_8 = dou_chanel[7][:, :, 1:]
final_c = np.concatenate((c_1, c_2, c_3, c_4, c_5, c_6, c_7, c_8), axis=2)
chanelfuse.append(final_c)
#
# dou_em = np.array(dou_em, dtype=np.float32)
# dou_em = np.array(dou_em[:, :, :])
# dou_em = np.expand_dims(dou_em, 2)
# double_edge.append(dou_em)
dou_em = np.array(dou_em, dtype=np.float32)
# 转化为无类别的GT 100 255 为边缘
dou_em = np.where(dou_em == 50, 0, dou_em)
dou_em = np.where(dou_em == 100, 1, dou_em)
dou_em = np.where(dou_em == 255, 1, dou_em)
dou_em = np.array(dou_em[:, :])
dou_em = np.expand_dims(dou_em, 2)
# dou_em = tran(dou_em)
double_edge.append(dou_em)
# 1转化为tensor
filenames.append(self.X_train[index])
images = np.asarray(images)
double_edge = np.asarray(double_edge)
chanel1 = np.asarray(chanel[0])
chanel2 = np.asarray(chanel[1])
chanel3 = np.asarray(chanel[2])
chanel4 = np.asarray(chanel[3])
chanel5 = np.asarray(chanel[4])
chanel6 = np.asarray(chanel[5])
chanel7 = np.asarray(chanel[6])
chanel8 = np.asarray(chanel[7])
chanelfuse = np.asarray(chanelfuse)
return images, edgemaps, double_edge, chanel1, chanel2, chanel3, chanel4, chanel5, chanel6, chanel7, chanel8, chanelfuse, edgemaps_4, edgemaps_8, edgemaps_16, filenames
def combine_augment_method(self, im, dou_em, relation_8_map, index):
"""
基于彪哥给的代码做修改
:param im:
:return:
"""
weight = im.size[0]
height = im.size[1]
if im.size[0] > 320 or im.size[1] > 320:
print('The input image path size is error: ', im.size[0],
im.size[1])
sys.exit('please check the code')
else:
# 1 决定图像是否加压缩
if random.randint(0, 20) == 1:
try:
mb = random.randint(30, 100)
path = compress_image(infile=self.X_train_or_test[index],
mb=mb)
im = Image.open(path)
except:
traceback.print_exc()
# 2 决定图像是否加模糊
if random.randint(0, 20) == 1:
m = random.randint(0, 5)
if m == 0:
r = random.randint(1, 3)
im = im.filter(ImageFilter.GaussianBlur(radius=r))
else:
pass
# 3 决定是否flip 旋转
random_aug = random.randint(0, 20)
try:
if random_aug < 5:
if random_aug == 0:
# 旋转90
im = im.transpose(Image.ROTATE_90)
dou_em = dou_em.transpose(Image.ROTATE_90)
for i in range(8):
relation_8_map[i] = relation_8_map[i].transpose(
Image.ROTATE_90)
elif random_aug == 1:
# 旋转180
im = im.transpose(Image.ROTATE_180)
dou_em = dou_em.transpose(Image.ROTATE_180)
for i in range(8):
relation_8_map[i] = relation_8_map[i].transpose(
Image.ROTATE_180)
elif random_aug == 2:
# 旋转270
im = im.transpose(Image.ROTATE_270)
dou_em = dou_em.transpose(Image.ROTATE_270)
for i in range(8):
relation_8_map[i] = relation_8_map[i].transpose(
Image.ROTATE_270)
elif random_aug == 3:
# 左右互换
im = im.transpose(Image.FLIP_LEFT_RIGHT)
dou_em = dou_em.transpose(Image.FLIP_LEFT_RIGHT)
for i in range(8):
relation_8_map[i] = relation_8_map[i].transpose(
Image.FLIP_LEFT_RIGHT)
elif random_aug == 4:
# 左右呼唤
im = im.transpose(Image.FLIP_TOP_BOTTOM)
dou_em = dou_em.transpose(Image.FLIP_TOP_BOTTOM)
for i in range(8):
relation_8_map[i] = relation_8_map[i].transpose(
Image.FLIP_TOP_BOTTOM)
except:
traceback.print_exc()
return im, dou_em
def get_batch(self, batch, train=True):
filenames = []
images = []
edgemaps = []
double_edge = []
edgemaps_4 = []
edgemaps_8 = []
edgemaps_16 = []
chanel1 = []
chanel2 = []
chanel3 = []
chanel4 = []
chanel5 = []
chanel6 = []
chanel7 = []
chanel8 = []
chanelfuse = []
for idx, b in enumerate(batch):
if train:
index = self.X_train.index(b)
im = Image.open(self.X_train[index])
dou_path = os.path.join(self.double_edge_file, self.Y_train[index].split('/')[-1])
dou_em = Image.open(dou_path)
# path = os.path.join(self.save_path_1_1,self.Y_train[index].split('/')[-1])
# chanel_1_1 = Image.open(path)
# # -1 0
# path = os.path.join(self.save_path_10,self.Y_train[index].split('/')[-1])
# chanel_10 = Image.open(path)
# # -1 1
# path = os.path.join(self.save_path_11,self.Y_train[index].split('/')[-1])
# chanel_11 = Image.open(path)
# # 0 -1
# path = os.path.join(self.save_path0_1,self.Y_train[index].split('/')[-1])
# chanel0_1 = Image.open(path)
# # 0 1
# path = os.path.join(self.save_path01,self.Y_train[index].split('/')[-1])
# chanel01 = Image.open(path)
# # 1 -1
# path = os.path.join(self.save_path1_1,self.Y_train[index].split('/')[-1])
# chanel1_1 = Image.open(path)
# # 1 0
# path = os.path.join(self.save_path10,self.Y_train[index].split('/')[-1])
# chanel10 = Image.open(path)
# # 1 1
# path = os.path.join(self.save_path11,self.Y_train[index].split('/')[-1])
# chanel11 = Image.open(path)
# 在这里获取8张图，从左上角按照顺时针顺序,返回的是一个长度为8的列表
relation_8_map = gen_8_map(dou_em)
for i in range(8):
relation_8_map[i] = Image.fromarray(np.uint8(relation_8_map[i]))
weight = im.size[0]
height = im.size[1]
if im.size[0] > 320 and im.size[1] > 320:
w_centry = weight // 2
h_centry = height // 2
if w_centry > 160:
range_w = random.randint(0, w_centry - 160)
else:
range_w = 0
if h_centry > 160:
range_h = random.randint(0, h_centry - 160)
else:
range_h = 0
# 决定图像是否加压缩
if random.randint(0, 20) == 1:
try:
mb = random.randint(30, 100)
path = compress_image(infile=self.X_train[index], mb=mb)
im = Image.open(path)
except:
traceback.print_exc()
# 决定图像是否加模糊
if random.randint(0, 20) == 1:
m = random.randint(0, 5)
if m == 0:
r = random.randint(1, 3)
im = im.filter(ImageFilter.GaussianBlur(radius=r))
else:
pass
if random.randint(0, 1) == 0:
im = im.crop((w_centry - 160 + range_w, h_centry - 160 + range_h, w_centry + 160 + range_w,
h_centry + 160 + range_h))
dou_em = dou_em.crop(
(w_centry - 160 + range_w, h_centry - 160 + range_h, w_centry + 160 + range_w,
h_centry + 160 + range_h))
for i in range(8):
relation_8_map[i] = relation_8_map[i].crop(
(w_centry - 160 + range_w, h_centry - 160 + range_h, w_centry + 160 + range_w,
h_centry + 160 + range_h))
else:
im = im.crop((w_centry - 160 - range_w, h_centry - 160 - range_h, w_centry + 160 - range_w,
h_centry + 160 - range_h))
dou_em = dou_em.crop(
(w_centry - 160 - range_w, h_centry - 160 - range_h, w_centry + 160 - range_w,
h_centry + 160 - range_h))
for i in range(8):
relation_8_map[i] = relation_8_map[i].crop(
(w_centry - 160 - range_w, h_centry - 160 - range_h, w_centry + 160 - range_w,
h_centry + 160 - range_h))
# 决定是否flip 旋转
random_aug = random.randint(0, 20)
try:
if random_aug < 5:
if random_aug == 0:
# 旋转90
im = im.transpose(Image.ROTATE_90)
dou_em = dou_em.transpose(Image.ROTATE_90)
for i in range(8):
relation_8_map[i] = relation_8_map[i].transpose(Image.ROTATE_90)
elif random_aug == 1:
# 旋转180
im = im.transpose(Image.ROTATE_180)
dou_em = dou_em.transpose(Image.ROTATE_180)
for i in range(8):
relation_8_map[i] = relation_8_map[i].transpose(Image.ROTATE_180)
elif random_aug == 2:
# 旋转270
im = im.transpose(Image.ROTATE_270)
dou_em = dou_em.transpose(Image.ROTATE_270)
for i in range(8):
relation_8_map[i] = relation_8_map[i].transpose(Image.ROTATE_270)
elif random_aug == 3:
# 左右呼唤
im = im.transpose(Image.FLIP_LEFT_RIGHT)
dou_em = dou_em.transpose(Image.FLIP_LEFT_RIGHT)
for i in range(8):
relation_8_map[i] = relation_8_map[i].transpose(Image.FLIP_LEFT_RIGHT)
elif random_aug == 4:
# 左右呼唤
im = im.transpose(Image.FLIP_TOP_BOTTOM)
dou_em = dou_em.transpose(Image.FLIP_TOP_BOTTOM)
for i in range(8):
relation_8_map[i] = relation_8_map[i].transpose(Image.FLIP_TOP_BOTTOM)
except:
traceback.print_exc()
else:
# 决定图像是否加压缩
if random.randint(0, 20) == 1:
# o_size = get_size(self.X_train[index])
try:
mb = random.randint(30, 100)
path = compress_image(infile=self.X_train[index], mb=mb)
im = Image.open(path)
except:
traceback.print_exc()
# 决定图像是否加模糊
if random.randint(0, 20) == 1:
m = random.randint(0, 5)
if m == 0:
r = random.randint(1, 3)
im = im.filter(ImageFilter.GaussianBlur(radius=r))
else:
pass
# x = random.randint(0, 200)
# y = random.randint(0, 200)
# bounds_x = random.randint(20, 100)
# bounds_y = random.randint(20, 100)
# bounds = (x, y, x + bounds_x, y + bounds_y)
# r = random.randint(1, 3)
# im = im.filter(MyGaussianBlur(radius=r, bounds=bounds))
im = im.crop((0, 0, self.image_height, self.image_width))
dou_em = dou_em.crop((0, 0, self.image_height, self.image_width))
for i in range(8):
relation_8_map[i] = relation_8_map[i].crop((0, 0, self.image_height, self.image_width))
# chanel_1_1 = chanel_1_1.crop((0, 0, self.image_height, self.image_width))
# chanel_10 = chanel_10.crop((0, 0, self.image_height, self.image_width))
# chanel_11 = chanel_11.crop((0, 0, self.image_height, self.image_width))
# chanel0_1 = chanel0_1.crop((0, 0, self.image_height, self.image_width))
# chanel01 = chanel01.crop((0, 0, self.image_height, self.image_width))
# chanel1_1 = chanel1_1.crop((0, 0, self.image_height, self.image_width))
# chanel10 = chanel10.crop((0, 0, self.image_height, self.image_width))
# chanel11 = chanel11.crop((0, 0, self.image_height, self.image_width))
# print(im)
im = np.array(im, dtype=np.float32)
im = im[..., ::-1] # RGB 2 BGR
# R = im[..., 0].mean()
# G = im[..., 1].mean()
# B = im[..., 2].mean()
# im[..., 0] -= R
# im[..., 1] -= G
# im[..., 2] -= B
# R=118.98194217348079 G=127.4061956623793 B=138.00865419127499
im[..., 0] -= 138.008
im[..., 1] -= 127.406
im[..., 2] -= 118.982
dou_chanel = [0 for i in range(8)]
chanel = [[] for i in range(8)]
for i in range(8):
dou_chanel[i] = np.array(relation_8_map[i], dtype=np.float32)
dou_chanel[i] = np.array(dou_chanel[i][:, :, 1:])
dou_chanel[i] = dou_chanel[i] / 255
chanel[i].append(dou_chanel[i])
c_1 = dou_chanel[0][:, :, 1:]
c_2 = dou_chanel[1][:, :, 1:]
c_3 = dou_chanel[2][:, :, 1:]
c_4 = dou_chanel[3][:, :, 1:]
c_5 = dou_chanel[4][:, :, 1:]
c_6 = dou_chanel[5][:, :, 1:]
c_7 = dou_chanel[6][:, :, 1:]
c_8 = dou_chanel[7][:, :, 1:]
final_c = np.concatenate((c_1, c_2, c_3, c_4, c_5, c_6, c_7, c_8), axis=2)
chanelfuse.append(final_c)
dou_em = np.array(dou_em, dtype=np.float32)
# dou_em = np.array(dou_em[:, :, :])
dou_em = np.array(dou_em[:, :])
double_edge.append(dou_em)
images.append(im)
# edgemaps.append(bin_em)
# edgemaps_4.append(bin_em4)
# edgemaps_8.append(bin_em8)
# edgemaps_16.append(bin_em16)
filenames.append(self.X_train[index])
else:
index = self.X_test.index(b)
im = Image.open(self.X_test[index])
# 决定图像是否加压缩
if random.randint(0, 20) == 1:
# o_size = get_size(self.X_train[index])
try:
mb = random.randint(30, 100)
path = compress_image(infile=self.X_test[index], mb=mb)
im = Image.open(path)
except:
traceback.print_exc()
# 决定图像是否加模糊
if random.randint(0, 20) == 1:
m = random.randint(0, 5)
if m == 0:
r = random.randint(1, 3)
im = im.filter(ImageFilter.GaussianBlur(radius=r))
else:
pass
# x = random.randint(0, 200)
# y = random.randint(0, 200)
# bounds_x = random.randint(20, 100)
# bounds_y = random.randint(20, 100)
# bounds = (x, y, x + bounds_x, y + bounds_y)
# r = random.randint(1, 3)
# im = im.filter(MyGaussianBlur(radius=r, bounds=bounds))
em = Image.open(self.Y_test[index])
# 这里不是dou_em了吗
# 在这里获取8张图，从左上角按照顺时针顺序，返回的是一个长度为8的列表
relation_8_map = gen_8_map(em)
for i in range(8):
relation_8_map[i] = Image.fromarray(relation_8_map[i])
# dou_path = os.path.join(self.double_edge_file, self.Y_test[index].split('/')[-1])
# dou_em = Image.open(dou_path)
#
# path = os.path.join(self.save_path_1_1, self.Y_test[index].split('/')[-1])
# chanel_1_1 = Image.open(path)
# # -1 0
# path = os.path.join(self.save_path_10, self.Y_test[index].split('/')[-1])
# chanel_10 = Image.open(path)
# # -1 1
# path = os.path.join(self.save_path_11, self.Y_test[index].split('/')[-1])
# chanel_11 = Image.open(path)
# # 0 -1
# path = os.path.join(self.save_path0_1, self.Y_test[index].split('/')[-1])
# chanel0_1 = Image.open(path)
# # 0 1
# path = os.path.join(self.save_path01, self.Y_test[index].split('/')[-1])
# chanel01 = Image.open(path)
# # 1 -1
# path = os.path.join(self.save_path1_1, self.Y_test[index].split('/')[-1])
# chanel1_1 = Image.open(path)
# # 1 0
# path = os.path.join(self.save_path10, self.Y_test[index].split('/')[-1])
# chanel10 = Image.open(path)
# # 1 1
# path = os.path.join(self.save_path11, self.Y_test[index].split('/')[-1])
# chanel11 = Image.open(path)
weight = im.size[0]
height = im.size[1]
if im.size[0] > 320 and im.size[1] > 320:
w_centry = weight // 2
h_centry = height // 2
if w_centry > 160:
range_w = random.randint(0, w_centry - 160)
else:
range_w = 0
if h_centry > 160:
range_h = random.randint(0, h_centry - 160)
else:
range_h = 0
if random.randint(0, 1) == 0:
im = im.crop((w_centry - 160 + range_w, h_centry - 160 + range_h, w_centry + 160 + range_w,
h_centry + 160 + range_h))
em = em.crop((w_centry - 160 + range_w, h_centry - 160 + range_h, w_centry + 160 + range_w,
h_centry + 160 + range_h))
dou_em = dou_em.crop(
(w_centry - 160 + range_w, h_centry - 160 + range_h, w_centry + 160 + range_w,
h_centry + 160 + range_h))
chanel_1_1 = chanel_1_1.crop(
(w_centry - 160 + range_w, h_centry - 160 + range_h, w_centry + 160 + range_w,
h_centry + 160 + range_h))
chanel_10 = chanel_10.crop(
(w_centry - 160 + range_w, h_centry - 160 + range_h, w_centry + 160 + range_w,
h_centry + 160 + range_h))
chanel_11 = chanel_11.crop(
(w_centry - 160 + range_w, h_centry - 160 + range_h, w_centry + 160 + range_w,
h_centry + 160 + range_h))
chanel0_1 = chanel0_1.crop(
(w_centry - 160 + range_w, h_centry - 160 + range_h, w_centry + 160 + range_w,
h_centry + 160 + range_h))
chanel01 = chanel01.crop(
(w_centry - 160 + range_w, h_centry - 160 + range_h, w_centry + 160 + range_w,
h_centry + 160 + range_h))
chanel1_1 = chanel1_1.crop(
(w_centry - 160 + range_w, h_centry - 160 + range_h, w_centry + 160 + range_w,
h_centry + 160 + range_h))
chanel10 = chanel10.crop(
(w_centry - 160 + range_w, h_centry - 160 + range_h, w_centry + 160 + range_w,
h_centry + 160 + range_h))
chanel11 = chanel11.crop(
(w_centry - 160 + range_w, h_centry - 160 + range_h, w_centry + 160 + range_w,
h_centry + 160 + range_h))
else:
im = im.crop((w_centry - 160 - range_w, h_centry - 160 - range_h, w_centry + 160 - range_w,
h_centry + 160 - range_h))
em = em.crop((w_centry - 160 - range_w, h_centry - 160 - range_h, w_centry + 160 - range_w,
h_centry + 160 - range_h))
dou_em = dou_em.crop(
(w_centry - 160 - range_w, h_centry - 160 - range_h, w_centry + 160 - range_w,
h_centry + 160 - range_h))
chanel_1_1 = chanel_1_1.crop(
(w_centry - 160 - range_w, h_centry - 160 - range_h, w_centry + 160 - range_w,
h_centry + 160 - range_h))
chanel_10 = chanel_10.crop(
(w_centry - 160 - range_w, h_centry - 160 - range_h, w_centry + 160 - range_w,
h_centry + 160 - range_h))
chanel_11 = chanel_11.crop(
(w_centry - 160 - range_w, h_centry - 160 - range_h, w_centry + 160 - range_w,
h_centry + 160 - range_h))
chanel0_1 = chanel0_1.crop(
(w_centry - 160 - range_w, h_centry - 160 - range_h, w_centry + 160 - range_w,
h_centry + 160 - range_h))
chanel01 = chanel01.crop(
(w_centry - 160 - range_w, h_centry - 160 - range_h, w_centry + 160 - range_w,
h_centry + 160 - range_h))
chanel1_1 = chanel1_1.crop(
(w_centry - 160 - range_w, h_centry - 160 - range_h, w_centry + 160 - range_w,
h_centry + 160 - range_h))
chanel10 = chanel10.crop(
(w_centry - 160 - range_w, h_centry - 160 - range_h, w_centry + 160 - range_w,
h_centry + 160 - range_h))
chanel11 = chanel11.crop(
(w_centry - 160 - range_w, h_centry - 160 - range_h, w_centry + 160 - range_w,
h_centry + 160 - range_h))
else:
im = im.crop((0, 0, self.image_height, self.image_width))
em = em.crop((0, 0, self.image_height, self.image_width))
dou_em = dou_em.crop((0, 0, self.image_height, self.image_width))
chanel_1_1 = chanel_1_1.crop((0, 0, self.image_height, self.image_width))
chanel_10 = chanel_10.crop((0, 0, self.image_height, self.image_width))
chanel_11 = chanel_11.crop((0, 0, self.image_height, self.image_width))
chanel0_1 = chanel0_1.crop((0, 0, self.image_height, self.image_width))
chanel01 = chanel01.crop((0, 0, self.image_height, self.image_width))
chanel1_1 = chanel1_1.crop((0, 0, self.image_height, self.image_width))
chanel10 = chanel10.crop((0, 0, self.image_height, self.image_width))
chanel11 = chanel11.crop((0, 0, self.image_height, self.image_width))
im = np.array(im, dtype=np.float32)
im = im[..., ::-1] # RGB 2 BGR
# R=118.98194217348079 G=127.4061956623793 B=138.00865419127499
im[..., 0] -= 138.008
im[..., 1] -= 127.406
im[..., 2] -= 118.982
# im[..., 0] -= 103.939
# im[..., 1] -= 116.779
# im[..., 2] -= 123.68
# R = im[..., 0].mean()
# G = im[..., 1].mean()
# B = im[..., 2].mean()
# im[..., 0] -= 103.939
# im[..., 1] -= 116.779
# im[..., 2] -= 123.68
# im[..., 0] -= R
# im[..., 1] -= G
# im[..., 2] -= B
# Labels needs to be 1 or 0 (edge pixel or not)
# or can use regression targets as done by the author
# https://github.com/s9xie/hed/blob/9e74dd710773d8d8a469ad905c76f4a7fa08f945/src/caffe/layers/image_labelmap_data_layer.cpp#L213
em_16 = em.resize((160, 160), Image.BICUBIC)
em_16 = np.array(em_16.convert('L'), dtype=np.float32)
em_16 = np.where(em_16 > 0, 255, 0)
# new_image = Image.fromarray(np.uint8(em_16))
em_8 = em.resize((80, 80), Image.BICUBIC)
em_8 = np.array(em_8.convert('L'), dtype=np.float32)
em_8 = np.where(em_8 > 0, 255, 0)
# new_image = Image.fromarray(np.uint8(em_8))
em_4 = em.resize((40, 40), Image.BICUBIC)
em_4 = np.array(em_4.convert('L'), dtype=np.float32)
em_4 = np.where(em_4 > 0, 255, 0)
em = np.array(em.convert('L'), dtype=np.float32)
if self.target_regression:
bin_em = em / 255.0
else:
bin_em = np.zeros_like(em)
bin_em[np.where(em)] = 1
bin_em = bin_em if bin_em.ndim == 2 else bin_em[:, :, 0]
bin_em = np.expand_dims(bin_em, 2)
if self.target_regression:
bin_em4 = em_4 / 255.0
else:
bin_em4 = np.zeros_like(em_4)
bin_em4[np.where(em_4)] = 1
bin_em4 = bin_em4 if bin_em4.ndim == 2 else bin_em4[:, :, 0]
bin_em4 = np.expand_dims(bin_em4, 2)
if self.target_regression:
bin_em8 = em_8 / 255.0
else:
bin_em8 = np.zeros_like(em_8)
bin_em8[np.where(em_8)] = 1
bin_em8 = bin_em8 if bin_em8.ndim == 2 else bin_em8[:, :, 0]
bin_em8 = np.expand_dims(bin_em8, 2)
if self.target_regression:
bin_em16 = em_16 / 255.0
else:
bin_em16 = np.zeros_like(em_16)
bin_em16[np.where(em_16)] = 1
bin_em16 = bin_em16 if bin_em16.ndim == 2 else bin_em16[:, :, 0]
bin_em16 = np.expand_dims(bin_em16, 2)
dou_em = np.array(dou_em, dtype=np.float32)
dou_em = np.array(dou_em[:, :, :])
# dou_em = np.expand_dims(dou_em, 2)
double_edge.append(dou_em)
images.append(im)
edgemaps.append(bin_em)
edgemaps_16.append(bin_em16)
edgemaps_8.append(bin_em8)
edgemaps_4.append(bin_em4)
filenames.append(self.X_test[index])
# _1_1
dou_chanel1 = np.array(chanel_1_1, dtype=np.float32)
dou_chanel1 = np.array(dou_chanel1[:, :, 1:])
dou_chanel1 = dou_chanel1 / 255
chanel1.append(dou_chanel1)
# chanel_1_1 = np.array(chanel_1_1.convert('L'), dtype=np.float32)
# if self.target_regression:
# bin_chanel_1_1 = chanel_1_1 / 255.0
# else:
# bin_chanel_1_1 = np.zeros_like(chanel_1_1)
# bin_chanel_1_1[np.where(chanel_1_1)] = 1
# bin_chanel_1_1 = bin_chanel_1_1 if bin_chanel_1_1.ndim == 2 else bin_chanel_1_1[:, :, 0]
# bin_chanel_1_1 = np.expand_dims(bin_chanel_1_1, 2)
# chanel1.append(bin_chanel_1_1)
# _1 0
dou_chanel2 = np.array(chanel_10, dtype=np.float32)
dou_chanel2 = np.array(dou_chanel2[:, :, 1:])
dou_chanel2 = dou_chanel2 / 255
chanel2.append(dou_chanel2)
# chanel_10 = np.array(chanel_10.convert('L'), dtype=np.float32)
# if self.target_regression:
# bin_chanel_10 = chanel_10 / 255.0
# else:
# bin_chanel_10 = np.zeros_like(chanel_10)
# bin_chanel_10[np.where(chanel_10)] = 1
# bin_chanel_10 = bin_chanel_10 if bin_chanel_10.ndim == 2 else bin_chanel_10[:, :, 0]
# bin_chanel_10 = np.expand_dims(bin_chanel_10, 2)
# chanel2.append(bin_chanel_10)
# _1 1
dou_chanel3 = np.array(chanel_11, dtype=np.float32)
dou_chanel3 = np.array(dou_chanel3[:, :, 1:])
dou_chanel3 = dou_chanel3 / 255
chanel3.append(dou_chanel3)
# chanel_11 = np.array(chanel_11.convert('L'), dtype=np.float32)
# if self.target_regression:
# bin_chanel_11 = chanel_11 / 255.0
# else:
# bin_chanel_11 = np.zeros_like(chanel_11)
# bin_chanel_11[np.where(chanel_11)] = 1
# bin_chanel_11 = bin_chanel_11 if bin_chanel_11.ndim == 2 else bin_chanel_11[:, :, 0]
# bin_chanel_11 = np.expand_dims(bin_chanel_11, 2)
# chanel3.append(bin_chanel_11)
# 0_1
dou_chanel4 = np.array(chanel0_1, dtype=np.float32)
dou_chanel4 = np.array(dou_chanel4[:, :, 1:])
dou_chanel4 = dou_chanel4 / 255
chanel4.append(dou_chanel4)
# chanel0_1 = np.array(chanel0_1.convert('L'), dtype=np.float32)
# if self.target_regression:
# bin_chanel0_1 = chanel0_1 / 255.0
# else:
# bin_chanel0_1 = np.zeros_like(chanel0_1)
# bin_chanel0_1[np.where(chanel0_1)] = 1
# bin_chanel0_1 = bin_chanel0_1 if bin_chanel0_1.ndim == 2 else bin_chanel0_1[:, :, 0]
# bin_chanel0_1 = np.expand_dims(bin_chanel0_1, 2)
# chanel4.append(bin_chanel0_1)
# 0 1
dou_chanel5 = np.array(chanel01, dtype=np.float32)
dou_chanel5 = np.array(dou_chanel5[:, :, 1:])
dou_chanel5 = dou_chanel5 / 255
chanel5.append(dou_chanel5)
# chanel01 = np.array(chanel01.convert('L'), dtype=np.float32)
# if self.target_regression:
# bin_chanel01 = chanel01 / 255.0
# else:
# bin_chanel01 = np.zeros_like(chanel01)
# bin_chanel01[np.where(chanel01)] = 1
# bin_chanel01 = bin_chanel01 if bin_chanel01.ndim == 2 else bin_chanel01[:, :, 0]
# bin_chanel01 = np.expand_dims(bin_chanel01, 2)
# chanel5.append(bin_chanel01)
# 1_1
dou_chanel6 = np.array(chanel1_1, dtype=np.float32)
dou_chanel6 = np.array(dou_chanel6[:, :, 1:])
dou_chanel6 = dou_chanel6 / 255
chanel6.append(dou_chanel6)
# chanel1_1 = np.array(chanel1_1.convert('L'), dtype=np.float32)
# if self.target_regression:
# bin_chanel1_1 = chanel1_1 / 255.0
# else:
# bin_chanel1_1 = np.zeros_like(chanel1_1)
# bin_chanel1_1[np.where(chanel1_1)] = 1
# bin_chanel1_1 = bin_chanel1_1 if bin_chanel1_1.ndim == 2 else bin_chanel1_1[:, :, 0]
# bin_chanel1_1 = np.expand_dims(bin_chanel1_1, 2)
# chanel6.append(bin_chanel1_1)
# 1 0
dou_chanel7 = np.array(chanel10, dtype=np.float32)
dou_chanel7 = np.array(dou_chanel7[:, :, 1:])
dou_chanel7 = dou_chanel7 / 255
chanel7.append(dou_chanel7)
# chanel10 = np.array(chanel10.convert('L'), dtype=np.float32)
# if self.target_regression:
# bin_chanel10 = chanel10 / 255.0
# else:
# bin_chanel10 = np.zeros_like(chanel10)
# bin_chanel10[np.where(chanel10)] = 1
# bin_chanel10 = bin_chanel10 if bin_chanel10.ndim == 2 else bin_chanel10[:, :, 0]
# bin_chanel10 = np.expand_dims(bin_chanel10, 2)
# chanel7.append(bin_chanel10)
# 1 1
dou_chanel8 = np.array(chanel11, dtype=np.float32)
dou_chanel8 = np.array(dou_chanel8[:, :, 1:])
dou_chanel8 = dou_chanel8 / 255
chanel8.append(dou_chanel8)
c_1 = dou_chanel1[:, :, 1:]
c_2 = dou_chanel2[:, :, 1:]
c_3 = dou_chanel3[:, :, 1:]
c_4 = dou_chanel4[:, :, 1:]
c_5 = dou_chanel5[:, :, 1:]
c_6 = dou_chanel6[:, :, 1:]
c_7 = dou_chanel7[:, :, 1:]
c_8 = dou_chanel8[:, :, 1:]
final_c = np.concatenate((c_1, c_2, c_3, c_4, c_5, c_6, c_7, c_8), axis=2)
# final_c = final_c.reshape(-1, 8)
# print(final_c.shape)
chanelfuse.append(final_c)
# chanel11 = np.array(chanel11.convert('L'), dtype=np.float32)
# if self.target_regression:
# bin_chanel11 = chanel11 / 255.0 1
# else:
# bin_chanel11 = np.zeros_like(chanel11)
# bin_chanel11[np.where(chanel11)] = 1
# bin_chanel11 = bin_chanel11 if bin_chanel11.ndim == 2 else bin_chanel11[:, :, 0]
# bin_chanel11 = np.expand_dims(bin_chanel11, 2)
# chanel8.append(bin_chanel11)
images = np.asarray(images)
# edgemaps = np.asarray(edgemaps)
# edgemaps_4 = np.asanyarray(edgemaps_4)
# edgemaps_8 = np.asanyarray(edgemaps_8)
# edgemaps_16 = np.asanyarray(edgemaps_16)
double_edge = np.asarray(double_edge)
chanel1 = np.asarray(chanel[0])
chanel2 = np.asarray(chanel[1])
chanel3 = np.asarray(chanel[2])
chanel4 = np.asarray(chanel[3])
chanel5 = np.asarray(chanel[4])
chanel6 = np.asarray(chanel[5])
chanel7 = np.asarray(chanel[6])
chanel8 = np.asarray(chanel[7])
chanelfuse = np.asarray(chanelfuse)
print('++++++++++++++++++++++++++++++++++++++')
print(type(chanel1))
print()
print(chanel1.shape)
return images, edgemaps, double_edge, chanel1, chanel2, chanel3, chanel4, chanel5, chanel6, chanel7, chanel8, chanelfuse, edgemaps_4, edgemaps_8, edgemaps_16, filenames