TensorFlow代码
数据处理：
输入为（344,288）的图片，首先进行归一化处理，然后下采样再上采样，得到输入input_，
def preprocess(path, scale=3):
"""
Preprocess single image file
(1) Read original image as YCbCr format (and grayscale as default)
(2) Normalize
(3) Apply image file with bicubic interpolation
Args:
path: file path of desired file
input_: image applied bicubic interpolation (low-resolution)
label_: image with original resolution (high-resolution)
"""
image = imread(path, is_grayscale=True)#(344,228)
label_ = modcrop(image, scale)#(342,228)modcrop得到scale的整数倍
# Must be normalized
image = image / 255.
label_ = label_ / 255.
input_ = scipy.ndimage.interpolation.zoom(label_, (1./scale), prefilter=False)#(114,76)
input_ = scipy.ndimage.interpolation.zoom(input_, (scale/1.), prefilter=False)#(342,228)
return input_, label_得到子图片
input_, label_ = preprocess(data[0], config.scale)
if len(input_.shape) == 3:
h, w, _ = input_.shape
else:
h, w = input_.shape#(342,228)
# Numbers of sub-images in height and width of image are needed to compute merge operation.
nx = ny = 0
for x in range(0, h-config.image_size+1, config.stride):#(0,310,21)nx=15,ny=10
nx += 1; ny = 0
for y in range(0, w-config.image_size+1, config.stride):#(0,196,21)
ny += 1
sub_input = input_[x:x+config.image_size, y:y+config.image_size] # [33 x 33]
sub_label = label_[x+int(padding):x+int(padding)+config.label_size, y+int(padding):y+int(padding)+config.label_size] # [21 x 21]
sub_input = sub_input.reshape([config.image_size, config.image_size, 1]) #(33,33,1)
sub_label = sub_label.reshape([config.label_size, config.label_size, 1]) #(21,21,1)
sub_input_sequence.append(sub_input)
sub_label_sequence.append(sub_label)
"""
len(sub_input_sequence) : the number of sub_input (33 x 33 x ch) in one image
(sub_input_sequence[0]).shape : (33, 33, 1)
"""
# Make list to numpy array. With this transform
arrdata = np.asarray(sub_input_sequence) # [?, 33, 33, 1]
arrlabel = np.asarray(sub_label_sequence) # [?, 21, 21, 1]
make_data(sess, arrdata, arrlabel)
if not config.is_train:
return nx, ny #15,10