52Hz/SRMNet_real_world_denoising

1import argparse
2import cv2
3import glob
4import numpy as np
5from collections import OrderedDict
6from skimage import img_as_ubyte
7import os
8import torch
9import requests
10from PIL import Image
11import torchvision.transforms.functional as TF
12import torch.nn.functional as F
13from natsort import natsorted
14from model.SRMNet import SRMNet
15
16def main():
17    parser = argparse.ArgumentParser(description='Demo Image Denoising')
18    parser.add_argument('--input_dir', default='test/', type=str, help='Input images')
19    parser.add_argument('--result_dir', default='result/', type=str, help='Directory for results')
20    parser.add_argument('--weights',
21                        default='experiments/pretrained_models/real_denoising_SRMNet.pth', type=str,
22                        help='Path to weights')
23
24    args = parser.parse_args()
25
26    inp_dir = args.input_dir
27    out_dir = args.result_dir
28
29    os.makedirs(out_dir, exist_ok=True)
30
31    files = natsorted(glob.glob(os.path.join(inp_dir, '*')))
32
33    if len(files) == 0:
34        raise Exception(f"No files found at {inp_dir}")
35
36    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
37
38    # Load corresponding models architecture and weights
39    model = SRMNet()
40    model = model.to(device)
41    model.eval()
42    load_checkpoint(model, args.weights)
43    
44
45    mul = 16
46    for file_ in files:
47        img = Image.open(file_).convert('RGB')
48        input_ = TF.to_tensor(img).unsqueeze(0).to(device)
49
50        # Pad the input if not_multiple_of 8
51        h, w = input_.shape[2], input_.shape[3]
52        H, W = ((h + mul) // mul) * mul, ((w + mul) // mul) * mul
53        padh = H - h if h % mul != 0 else 0
54        padw = W - w if w % mul != 0 else 0
55        input_ = F.pad(input_, (0, padw, 0, padh), 'reflect')
56        with torch.no_grad():
57            restored = model(input_)
58
59        restored = torch.clamp(restored, 0, 1)
60        restored = restored[:, :, :h, :w]
61        restored = restored.permute(0, 2, 3, 1).cpu().detach().numpy()
62        restored = img_as_ubyte(restored[0])
63
64        f = os.path.splitext(os.path.split(file_)[-1])[0]
65        save_img((os.path.join(out_dir, f + '.png')), restored)
66
67
68def save_img(filepath, img):
69    cv2.imwrite(filepath, cv2.cvtColor(img, cv2.COLOR_RGB2BGR))
70
71
72def load_checkpoint(model, weights):
73    checkpoint = torch.load(weights, map_location=torch.device('cpu'))
74    try:
75        model.load_state_dict(checkpoint["state_dict"])
76    except:
77        state_dict = checkpoint["state_dict"]
78        new_state_dict = OrderedDict()
79        for k, v in state_dict.items():
80            name = k[7:]  # remove `module.`
81            new_state_dict[name] = v
82        model.load_state_dict(new_state_dict)
83
84
85if __name__ == '__main__':
86    main()