Image-Denoising-DNN/gen_data.py at master · MerlinPCarson/Image-Denoising-DNN · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
import os
import sys
import numpy as np
import argparse
import cv2 as cv
import h5py
from glob import glob
from tqdm import tqdm

#import matplotlib.pyplot as plt

def augment_img(img, aug_type):
    if aug_type == 'mirror':
        return np.fliplr(img)
    if aug_type == 'flip':
        return np.flipud(img)
    if aug_type == 'rotL':
        return np.rot90(img, k=1)
    if aug_type == 'rotR':
        return np.rot90(img, k=-1)


def image_augment(img, num_augs):
    # save as channel first
    data_aug = np.empty((num_augs+1,img.shape[2], img.shape[0], img.shape[1]))
    aug_types = np.array(['mirror', 'flip', 'rotL', 'rotR'])
    np.random.shuffle(aug_types)
    data_aug[0] = np.einsum('ijk->kij', img.astype(np.float32))
    for i in range(num_augs):
        data_aug[i+1] = np.einsum('ijk->kij', augment_img(img, aug_types[i]).astype(np.float32))

    return data_aug


def generate_data(train_path, val_path, patch_size, stride, scaling_factors, num_augments, num_channels):
    #num_channels = 3
    print(f'[Data Generation] Creating training data from {train_path} with {num_channels} channels')
    num_train = 0
    h5f = h5py.File('train.h5', 'w')
    num_train = 0
    for f in tqdm(sorted(glob(os.path.join(train_path, '*.png')))):
        #print(f'{num_train+1}: Preprocessing {f}')
        img = cv.imread(f)
        height, width, ch = img.shape

        for scale in scaling_factors:
            img_scaled = cv.resize(img, (int(height*scale), int(width*scale)), interpolation=cv.INTER_CUBIC)
            img_scaled = np.array(img_scaled[:,:,:num_channels].reshape((img_scaled.shape[0],img_scaled.shape[1],num_channels))/255)
            patches = get_image_patches(img_scaled, patch_size, stride)
            #print(f'  scaling: {scale}, num patches: {patches.shape[0]}')
            for patch_num in range(patches.shape[0]):
                data_aug = image_augment(patches[patch_num], num_augments)
                for aug in range(data_aug.shape[0]):
                    h5f.create_dataset(str(num_train), data=data_aug[aug])
                    num_train += 1

    h5f.close()

    print(f'[Data Generation] Creating validation data from {val_path}')
    num_val = 0
    h5f = h5py.File('val.h5', 'w')
    for f in tqdm(sorted(glob(os.path.join(val_path, '*.png')))):
        #print(f'Preprocessing {f}')
        img = cv.imread(f)
        img = np.array(img[:,:,:num_channels].reshape((img.shape[0],img.shape[1],num_channels))/255)
        patches = get_image_patches(img, patch_size, stride)
        for patch_num in range(patches.shape[0]):
            # channels first
            patch = np.einsum('ijk->kij', patches[patch_num].astype(np.float32))
            h5f.create_dataset(str(num_val), data=patch)
            num_val += 1
    h5f.close()

    print(f'Number of training examples {num_train}')
    print(f'Number of validation examples {num_val}')


def get_image_patches(img, patch_size, stride):
    win_row_end = img.shape[0] - patch_size
    win_col_end = img.shape[1] - patch_size
    num_patches_rows = int((img.shape[0]-patch_size)/stride + 1)
    num_patches_cols = int((img.shape[1]-patch_size)/stride + 1)
    num_chs = int(img.shape[2])
    total_patches = int(num_patches_rows * num_patches_cols)

    patches = np.zeros((total_patches, patch_size, patch_size, num_chs), dtype=float)

    rows = np.arange(0,win_row_end+1, stride)
    cols = np.arange(0,win_col_end+1, stride)
    patch_num = 0
    for row in rows:
        for col in cols:
            patch = img[row:row+patch_size, col:col+patch_size,:]
            patches[patch_num,:,:,:] = patch
            patch_num += 1

    return patches

def main():

    script_dir = os.path.dirname(os.path.realpath(__file__))

    parser = argparse.ArgumentParser(description="DnCNN-data generation")
    parser.add_argument("--train_path", type=str, default='data/train_color/train', help='root directory for training data')
    parser.add_argument("--val_path", type=str, default='data/train_color/val', help='root directory for validation data')
    parser.add_argument("--patch_size", type=int, default=50, help="image patch size to train on")
    parser.add_argument("--stride", type=int, default=10, help="image patch stride")
    parser.add_argument("--scaling_factors", type=str, default='1,.6,.4,.2', help="image scaling")
    parser.add_argument("--num_augments", type=int, default=0, help="number of data augmentations per patch")
    parser.add_argument("--num_channels", type=int, default=3, help="number of channels (bw=1, color=3)")
    args = parser.parse_args()

    train_path = os.path.join(script_dir, args.train_path)
    val_path = os.path.join(script_dir, args.val_path)
    patch_size = args.patch_size
    stride = args.stride
    scaling_factors = [float(scale) for scale in args.scaling_factors.split(',')]
    num_augments = args.num_augments

    print(f'[args] training data: {train_path}')
    print(f'[args] validation data: {val_path}')
    print(f'[args] patch size: {patch_size}, stride: {stride}')
    print(f'[args] scaling factors: {scaling_factors}')
    print(f'[args] number of augmentations: {num_augments}')

    generate_data(train_path, val_path, patch_size, stride, scaling_factors, num_augments, args.num_channels)

    return 0

if __name__ == '__main__':
    sys.exit(main())