BasicSR

1
from torch.utils import data as data
2
from torchvision.transforms.functional import normalize
3

4
from basicsr.data.data_util import paired_paths_from_folder, paired_paths_from_lmdb, paired_paths_from_meta_info_file
5
from basicsr.data.transforms import augment, paired_random_crop
6
from basicsr.utils import FileClient, bgr2ycbcr, imfrombytes, img2tensor
7
from basicsr.utils.registry import DATASET_REGISTRY
8

9

10
@DATASET_REGISTRY.register()
11
class PairedImageDataset(data.Dataset):
12
    """Paired image dataset for image restoration.
13

14
    Read LQ (Low Quality, e.g. LR (Low Resolution), blurry, noisy, etc) and GT image pairs.
15

16
    There are three modes:
17

18
    1. **lmdb**: Use lmdb files. If opt['io_backend'] == lmdb.
19
    2. **meta_info_file**: Use meta information file to generate paths. \
20
        If opt['io_backend'] != lmdb and opt['meta_info_file'] is not None.
21
    3. **folder**: Scan folders to generate paths. The rest.
22

23
    Args:
24
        opt (dict): Config for train datasets. It contains the following keys:
25
        dataroot_gt (str): Data root path for gt.
26
        dataroot_lq (str): Data root path for lq.
27
        meta_info_file (str): Path for meta information file.
28
        io_backend (dict): IO backend type and other kwarg.
29
        filename_tmpl (str): Template for each filename. Note that the template excludes the file extension.
30
            Default: '{}'.
31
        gt_size (int): Cropped patched size for gt patches.
32
        use_hflip (bool): Use horizontal flips.
33
        use_rot (bool): Use rotation (use vertical flip and transposing h and w for implementation).
34
        scale (bool): Scale, which will be added automatically.
35
        phase (str): 'train' or 'val'.
36
    """
37

38
    def __init__(self, opt):
39
        super(PairedImageDataset, self).__init__()
40
        self.opt = opt
41
        # file client (io backend)
42
        self.file_client = None
43
        self.io_backend_opt = opt['io_backend']
44
        self.mean = opt['mean'] if 'mean' in opt else None
45
        self.std = opt['std'] if 'std' in opt else None
46

47
        self.gt_folder, self.lq_folder = opt['dataroot_gt'], opt['dataroot_lq']
48
        if 'filename_tmpl' in opt:
49
            self.filename_tmpl = opt['filename_tmpl']
50
        else:
51
            self.filename_tmpl = '{}'
52

53
        if self.io_backend_opt['type'] == 'lmdb':
54
            self.io_backend_opt['db_paths'] = [self.lq_folder, self.gt_folder]
55
            self.io_backend_opt['client_keys'] = ['lq', 'gt']
56
            self.paths = paired_paths_from_lmdb([self.lq_folder, self.gt_folder], ['lq', 'gt'])
57
        elif 'meta_info_file' in self.opt and self.opt['meta_info_file'] is not None:
58
            self.paths = paired_paths_from_meta_info_file([self.lq_folder, self.gt_folder], ['lq', 'gt'],
59
                                                          self.opt['meta_info_file'], self.filename_tmpl)
60
        else:
61
            self.paths = paired_paths_from_folder([self.lq_folder, self.gt_folder], ['lq', 'gt'], self.filename_tmpl)
62

63
    def __getitem__(self, index):
64
        if self.file_client is None:
65
            self.file_client = FileClient(self.io_backend_opt.pop('type'), **self.io_backend_opt)
66

67
        scale = self.opt['scale']
68

69
        # Load gt and lq images. Dimension order: HWC; channel order: BGR;
70
        # image range: [0, 1], float32.
71
        gt_path = self.paths[index]['gt_path']
72
        img_bytes = self.file_client.get(gt_path, 'gt')
73
        img_gt = imfrombytes(img_bytes, float32=True)
74
        lq_path = self.paths[index]['lq_path']
75
        img_bytes = self.file_client.get(lq_path, 'lq')
76
        img_lq = imfrombytes(img_bytes, float32=True)
77

78
        # augmentation for training
79
        if self.opt['phase'] == 'train':
80
            gt_size = self.opt['gt_size']
81
            # random crop
82
            img_gt, img_lq = paired_random_crop(img_gt, img_lq, gt_size, scale, gt_path)
83
            # flip, rotation
84
            img_gt, img_lq = augment([img_gt, img_lq], self.opt['use_hflip'], self.opt['use_rot'])
85

86
        # color space transform
87
        if 'color' in self.opt and self.opt['color'] == 'y':
88
            img_gt = bgr2ycbcr(img_gt, y_only=True)[..., None]
89
            img_lq = bgr2ycbcr(img_lq, y_only=True)[..., None]
90

91
        # crop the unmatched GT images during validation or testing, especially for SR benchmark datasets
92
        # TODO: It is better to update the datasets, rather than force to crop
93
        if self.opt['phase'] != 'train':
94
            img_gt = img_gt[0:img_lq.shape[0] * scale, 0:img_lq.shape[1] * scale, :]
95

96
        # BGR to RGB, HWC to CHW, numpy to tensor
97
        img_gt, img_lq = img2tensor([img_gt, img_lq], bgr2rgb=True, float32=True)
98
        # normalize
99
        if self.mean is not None or self.std is not None:
100
            normalize(img_lq, self.mean, self.std, inplace=True)
101
            normalize(img_gt, self.mean, self.std, inplace=True)
102

103
        return {'lq': img_lq, 'gt': img_gt, 'lq_path': lq_path, 'gt_path': gt_path}
104

105
    def __len__(self):
106
        return len(self.paths)
107