File size: 12,341 Bytes
80e30a3
 
 
 
 
 
bbd0a76
80e30a3
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
import argparse
import cv2
import glob
import numpy as np
from collections import OrderedDict
import os
os.system("pip install timm")
import torch
import requests

from models.network_swinir import SwinIR as net
from utils import util_calculate_psnr_ssim as util


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument('--task', type=str, default='color_dn', help='classical_sr, lightweight_sr, real_sr, '
                                                                     'gray_dn, color_dn, jpeg_car')
    parser.add_argument('--scale', type=int, default=1, help='scale factor: 1, 2, 3, 4, 8') # 1 for dn and jpeg car
    parser.add_argument('--noise', type=int, default=15, help='noise level: 15, 25, 50')
    parser.add_argument('--jpeg', type=int, default=40, help='scale factor: 10, 20, 30, 40')
    parser.add_argument('--training_patch_size', type=int, default=128, help='patch size used in training SwinIR. '
                                       'Just used to differentiate two different settings in Table 2 of the paper. '
                                       'Images are NOT tested patch by patch.')
    parser.add_argument('--large_model', action='store_true', help='use large model, only provided for real image sr')
    parser.add_argument('--model_path', type=str,
                        default='model_zoo/swinir/001_classicalSR_DIV2K_s48w8_SwinIR-M_x2.pth')
    parser.add_argument('--folder_lq', type=str, default=None, help='input low-quality test image folder')
    parser.add_argument('--folder_gt', type=str, default=None, help='input ground-truth test image folder')
    args = parser.parse_args()

    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    # set up model
    if os.path.exists(args.model_path):
        print(f'loading model from {args.model_path}')
    else:
        os.makedirs(os.path.dirname(args.model_path), exist_ok=True)
        url = 'https://github.com/JingyunLiang/SwinIR/releases/download/v0.0/{}'.format(os.path.basename(args.model_path))
        r = requests.get(url, allow_redirects=True)
        print(f'downloading model {args.model_path}')
        open(args.model_path, 'wb').write(r.content)

    model = define_model(args)
    model.eval()
    model = model.to(device)

    # setup folder and path
    folder, save_dir, border, window_size = setup(args)
    os.makedirs(save_dir, exist_ok=True)
    test_results = OrderedDict()
    test_results['psnr'] = []
    test_results['ssim'] = []
    test_results['psnr_y'] = []
    test_results['ssim_y'] = []
    test_results['psnr_b'] = []
    psnr, ssim, psnr_y, ssim_y, psnr_b = 0, 0, 0, 0, 0

    for idx, path in enumerate(sorted(glob.glob(os.path.join(folder, '*')))):
        # read image
        imgname, img_lq, img_gt = get_image_pair(args, path)  # image to HWC-BGR, float32
        img_lq = np.transpose(img_lq if img_lq.shape[2] == 1 else img_lq[:, :, [2, 1, 0]], (2, 0, 1))  # HCW-BGR to CHW-RGB
        img_lq = torch.from_numpy(img_lq).float().unsqueeze(0).to(device)  # CHW-RGB to NCHW-RGB

        # inference
        with torch.no_grad():
            # pad input image to be a multiple of window_size
            _, _, h_old, w_old = img_lq.size()
            h_pad = (h_old // window_size + 1) * window_size - h_old
            w_pad = (w_old // window_size + 1) * window_size - w_old
            img_lq = torch.cat([img_lq, torch.flip(img_lq, [2])], 2)[:, :, :h_old + h_pad, :]
            img_lq = torch.cat([img_lq, torch.flip(img_lq, [3])], 3)[:, :, :, :w_old + w_pad]
            output = model(img_lq)
            output = output[..., :h_old * args.scale, :w_old * args.scale]

        # save image
        output = output.data.squeeze().float().cpu().clamp_(0, 1).numpy()
        if output.ndim == 3:
            output = np.transpose(output[[2, 1, 0], :, :], (1, 2, 0))  # CHW-RGB to HCW-BGR
        output = (output * 255.0).round().astype(np.uint8)  # float32 to uint8
        cv2.imwrite(f'{save_dir}/{imgname}_SwinIR.png', output)

        # evaluate psnr/ssim/psnr_b
        if img_gt is not None:
            img_gt = (img_gt * 255.0).round().astype(np.uint8)  # float32 to uint8
            img_gt = img_gt[:h_old * args.scale, :w_old * args.scale, ...]  # crop gt
            img_gt = np.squeeze(img_gt)

            psnr = util.calculate_psnr(output, img_gt, crop_border=border)
            ssim = util.calculate_ssim(output, img_gt, crop_border=border)
            test_results['psnr'].append(psnr)
            test_results['ssim'].append(ssim)
            if img_gt.ndim == 3:  # RGB image
                psnr_y = util.calculate_psnr(output, img_gt, crop_border=border, test_y_channel=True)
                ssim_y = util.calculate_ssim(output, img_gt, crop_border=border, test_y_channel=True)
                test_results['psnr_y'].append(psnr_y)
                test_results['ssim_y'].append(ssim_y)
            if args.task in ['jpeg_car']:
                psnr_b = util.calculate_psnrb(output, img_gt, crop_border=border, test_y_channel=True)
                test_results['psnr_b'].append(psnr_b)
            print('Testing {:d} {:20s} - PSNR: {:.2f} dB; SSIM: {:.4f}; '
                  'PSNR_Y: {:.2f} dB; SSIM_Y: {:.4f}; '
                  'PSNR_B: {:.2f} dB.'.
                  format(idx, imgname, psnr, ssim, psnr_y, ssim_y, psnr_b))
        else:
            print('Testing {:d} {:20s}'.format(idx, imgname))

    # summarize psnr/ssim
    if img_gt is not None:
        ave_psnr = sum(test_results['psnr']) / len(test_results['psnr'])
        ave_ssim = sum(test_results['ssim']) / len(test_results['ssim'])
        print('\n{} \n-- Average PSNR/SSIM(RGB): {:.2f} dB; {:.4f}'.format(save_dir, ave_psnr, ave_ssim))
        if img_gt.ndim == 3:
            ave_psnr_y = sum(test_results['psnr_y']) / len(test_results['psnr_y'])
            ave_ssim_y = sum(test_results['ssim_y']) / len(test_results['ssim_y'])
            print('-- Average PSNR_Y/SSIM_Y: {:.2f} dB; {:.4f}'.format(ave_psnr_y, ave_ssim_y))
        if args.task in ['jpeg_car']:
            ave_psnr_b = sum(test_results['psnr_b']) / len(test_results['psnr_b'])
            print('-- Average PSNR_B: {:.2f} dB'.format(ave_psnr_b))


def define_model(args):
    # 001 classical image sr
    if args.task == 'classical_sr':
        model = net(upscale=args.scale, in_chans=3, img_size=args.training_patch_size, window_size=8,
                    img_range=1., depths=[6, 6, 6, 6, 6, 6], embed_dim=180, num_heads=[6, 6, 6, 6, 6, 6],
                    mlp_ratio=2, upsampler='pixelshuffle', resi_connection='1conv')
        param_key_g = 'params'

    # 002 lightweight image sr
    # use 'pixelshuffledirect' to save parameters
    elif args.task == 'lightweight_sr':
        model = net(upscale=args.scale, in_chans=3, img_size=64, window_size=8,
                    img_range=1., depths=[6, 6, 6, 6], embed_dim=60, num_heads=[6, 6, 6, 6],
                    mlp_ratio=2, upsampler='pixelshuffledirect', resi_connection='1conv')
        param_key_g = 'params'

    # 003 real-world image sr
    elif args.task == 'real_sr':
        if not args.large_model:
            # use 'nearest+conv' to avoid block artifacts
            model = net(upscale=4, in_chans=3, img_size=64, window_size=8,
                        img_range=1., depths=[6, 6, 6, 6, 6, 6], embed_dim=180, num_heads=[6, 6, 6, 6, 6, 6],
                        mlp_ratio=2, upsampler='nearest+conv', resi_connection='1conv')
        else:
            # larger model size; use '3conv' to save parameters and memory; use ema for GAN training
            model = net(upscale=4, in_chans=3, img_size=64, window_size=8,
                        img_range=1., depths=[6, 6, 6, 6, 6, 6, 6, 6, 6], embed_dim=248,
                        num_heads=[8, 8, 8, 8, 8, 8, 8, 8, 8],
                        mlp_ratio=2, upsampler='nearest+conv', resi_connection='3conv')
        param_key_g = 'params_ema'

    # 004 grayscale image denoising
    elif args.task == 'gray_dn':
        model = net(upscale=1, in_chans=1, img_size=128, window_size=8,
                    img_range=1., depths=[6, 6, 6, 6, 6, 6], embed_dim=180, num_heads=[6, 6, 6, 6, 6, 6],
                    mlp_ratio=2, upsampler='', resi_connection='1conv')
        param_key_g = 'params'

    # 005 color image denoising
    elif args.task == 'color_dn':
        model = net(upscale=1, in_chans=3, img_size=128, window_size=8,
                    img_range=1., depths=[6, 6, 6, 6, 6, 6], embed_dim=180, num_heads=[6, 6, 6, 6, 6, 6],
                    mlp_ratio=2, upsampler='', resi_connection='1conv')
        param_key_g = 'params'

    # 006 JPEG compression artifact reduction
    # use window_size=7 because JPEG encoding uses 8x8; use img_range=255 because it's sligtly better than 1
    elif args.task == 'jpeg_car':
        model = net(upscale=1, in_chans=1, img_size=126, window_size=7,
                    img_range=255., depths=[6, 6, 6, 6, 6, 6], embed_dim=180, num_heads=[6, 6, 6, 6, 6, 6],
                    mlp_ratio=2, upsampler='', resi_connection='1conv')
        param_key_g = 'params'
    
    pretrained_model = torch.load(args.model_path)
    model.load_state_dict(pretrained_model[param_key_g] if param_key_g in pretrained_model.keys() else pretrained_model, strict=True)
        
    return model


def setup(args):
    # 001 classical image sr/ 002 lightweight image sr
    if args.task in ['classical_sr', 'lightweight_sr']:
        save_dir = f'results/swinir_{args.task}_x{args.scale}'
        folder = args.folder_gt
        border = args.scale
        window_size = 8

    # 003 real-world image sr
    elif args.task in ['real_sr']:
        save_dir = f'results/swinir_{args.task}_x{args.scale}'
        folder = args.folder_lq
        border = 0
        window_size = 8

    # 004 grayscale image denoising/ 005 color image denoising
    elif args.task in ['gray_dn', 'color_dn']:
        save_dir = f'results/swinir_{args.task}_noise{args.noise}'
        folder = args.folder_gt
        border = 0
        window_size = 8

    # 006 JPEG compression artifact reduction
    elif args.task in ['jpeg_car']:
        save_dir = f'results/swinir_{args.task}_jpeg{args.jpeg}'
        folder = args.folder_gt
        border = 0
        window_size = 7

    return folder, save_dir, border, window_size


def get_image_pair(args, path):
    (imgname, imgext) = os.path.splitext(os.path.basename(path))

    # 001 classical image sr/ 002 lightweight image sr (load lq-gt image pairs)
    if args.task in ['classical_sr', 'lightweight_sr']:
        img_gt = cv2.imread(path, cv2.IMREAD_COLOR).astype(np.float32) / 255.
        img_lq = cv2.imread(f'{args.folder_lq}/{imgname}x{args.scale}{imgext}', cv2.IMREAD_COLOR).astype(
            np.float32) / 255.

    # 003 real-world image sr (load lq image only)
    elif args.task in ['real_sr']:
        img_gt = None
        img_lq = cv2.imread(path, cv2.IMREAD_COLOR).astype(np.float32) / 255.

    # 004 grayscale image denoising (load gt image and generate lq image on-the-fly)
    elif args.task in ['gray_dn']:
        img_gt = cv2.imread(path, cv2.IMREAD_GRAYSCALE).astype(np.float32) / 255.
        np.random.seed(seed=0)
        img_lq = img_gt + np.random.normal(0, args.noise / 255., img_gt.shape)
        img_gt = np.expand_dims(img_gt, axis=2)
        img_lq = np.expand_dims(img_lq, axis=2)

    # 005 color image denoising (load gt image and generate lq image on-the-fly)
    elif args.task in ['color_dn']:
        img_gt = cv2.imread(path, cv2.IMREAD_COLOR).astype(np.float32) / 255.
        np.random.seed(seed=0)
        img_lq = img_gt + np.random.normal(0, args.noise / 255., img_gt.shape)

    # 006 JPEG compression artifact reduction (load gt image and generate lq image on-the-fly)
    elif args.task in ['jpeg_car']:
        img_gt = cv2.imread(path, cv2.IMREAD_UNCHANGED)
        if img_gt.ndim != 2:
            img_gt = util.rgb2ycbcr(cv2.cvtColor(img_gt, cv2.COLOR_BGR2RGB), y_only=True)
        result, encimg = cv2.imencode('.jpg', img_gt, [int(cv2.IMWRITE_JPEG_QUALITY), args.jpeg])
        img_lq = cv2.imdecode(encimg, 0)
        img_gt = np.expand_dims(img_gt, axis=2).astype(np.float32) / 255.
        img_lq = np.expand_dims(img_lq, axis=2).astype(np.float32) / 255.

    return imgname, img_lq, img_gt


if __name__ == '__main__':
    main()