scan-DocScanner/inference.py

from model import DocScanner
from seg import U2NETP

import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
import cv2
import os
from PIL import Image
import argparse
import warnings

warnings.filterwarnings('ignore')


class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.msk = U2NETP(3, 1)
        self.bm = DocScanner()

    def forward(self, x):
        msk, *_ = self.msk(x)
        msk = (msk > 0.5).float()
        x = msk * x

        bm = self.bm(x, iters=12, test_mode=True)
        bm = (2 * (bm / 286.8) - 1) * 0.99
        return bm


def load_model(model, path, strip_prefix=False):
    """Универсальная загрузка весов"""
    if not path or not os.path.exists(path):
        raise FileNotFoundError(f"Model not found: {path}")
    
    model_dict = model.state_dict()
    pretrained_dict = torch.load(path, map_location='cuda:0')
    
    if strip_prefix:
        # Для seg модели ключи имеют префикс 'module.' (6 символов)
        pretrained_dict = {k[6:]: v for k, v in pretrained_dict.items() if k[6:] in model_dict}
    else:
        pretrained_dict = {k: v for k, v in pretrained_dict.items() if k in model_dict}
        
    model_dict.update(pretrained_dict)
    model.load_state_dict(model_dict)
    return model


def main():
    parser = argparse.ArgumentParser(description='DocScanner Single File Inference')
    parser.add_argument('-i', '--input', required=True, help='Input image path')
    parser.add_argument('-o', '--output', required=True, help='Output image path')
    parser.add_argument('--seg_model', default='./model_pretrained/seg.pth')
    parser.add_argument('--rec_model', default='./model_pretrained/DocScanner-L.pth')
    opt = parser.parse_args()

    # 1. Init & Load Models
    net = Net().cuda()
    net.msk = load_model(net.msk, opt.seg_model, strip_prefix=True)
    net.bm = load_model(net.bm, opt.rec_model, strip_prefix=False)
    net.eval()

    # 2. Preprocess
    im_ori = np.array(Image.open(opt.input))[:, :, :3] / 255.0
    h, w, _ = im_ori.shape
    
    im = cv2.resize(im_ori, (288, 288))
    im = torch.from_numpy(im.transpose(2, 0, 1)).float().unsqueeze(0)

    # 3. Inference
    with torch.no_grad():
        bm = net(im.cuda()).cpu()

    # 4. Postprocess & Warp
    bm0 = cv2.blur(cv2.resize(bm[0, 0].numpy(), (w, h)), (3, 3))
    bm1 = cv2.blur(cv2.resize(bm[0, 1].numpy(), (w, h)), (3, 3))
    
    lbl = torch.from_numpy(np.stack([bm0, bm1], axis=2)).unsqueeze(0)
    inp_tensor = torch.from_numpy(im_ori).permute(2, 0, 1).unsqueeze(0).float()
    
    out = F.grid_sample(inp_tensor, lbl, align_corners=True)
    
    # 5. Save
    result = ((out[0] * 255).permute(1, 2, 0).numpy()[:, :, ::-1]).astype(np.uint8)
    
    out_dir = os.path.dirname(opt.output)
    if out_dir and not os.path.exists(out_dir):
        os.makedirs(out_dir)
        
    cv2.imwrite(opt.output, result)
    print(f"[OK] Saved: {opt.output}")


if __name__ == "__main__":
    main()