scan-DocScanner/inference.py

import os, sys, argparse, warnings
import torch, torch.nn as nn, torch.nn.functional as F
import numpy as np, cv2
from PIL import Image
from model import DocScanner
from seg import U2NETP

warnings.filterwarnings('ignore')

# Запоминаем CWD пользователя ДО смены директории
USER_CWD = os.getcwd()
SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
os.chdir(SCRIPT_DIR)


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

    def forward(self, x):
        msk, *_ = self.msk(x)
        bm = self.bm((msk > 0.5).float() * x, iters=12, test_mode=True)
        return (2 * (bm / 286.8) - 1) * 0.99


def load_model(model, path, strip_prefix=False):
    state_dict = model.state_dict()
    pretrained = torch.load(path, map_location='cuda:0')
    if strip_prefix:
        pretrained = {k[6:]: v for k, v in pretrained.items() if k[6:] in state_dict}
    else:
        pretrained = {k: v for k, v in pretrained.items() if k in state_dict}
    state_dict.update(pretrained)
    model.load_state_dict(state_dict)
    return model


def resolve_user_path(path):
    """Если путь относительный — считаем его относительно CWD пользователя, не скрипта"""
    if os.path.isabs(path):
        return path
    return os.path.join(USER_CWD, path)


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument('-i', '--input', required=True)
    parser.add_argument('-o', '--output', required=True)
    opt = parser.parse_args()

    input_path = resolve_user_path(opt.input)
    output_path = resolve_user_path(opt.output)

    net = Net().cuda().eval()
    load_model(net.msk, f'{SCRIPT_DIR}/model_pretrained/seg.pth', strip_prefix=True)
    load_model(net.bm, f'{SCRIPT_DIR}/model_pretrained/DocScanner-L.pth')

    image = np.array(Image.open(input_path))[:, :, :3] / 255.0
    height, width = image.shape[:2]
    tensor = torch.from_numpy(cv2.resize(image, (288, 288)).transpose(2, 0, 1)).float().unsqueeze(0)

    with torch.no_grad():
        bm = net(tensor.cuda()).cpu()

    flow = torch.from_numpy(np.stack([
        cv2.blur(cv2.resize(bm[0, 0].numpy(), (width, height)), (3, 3)),
        cv2.blur(cv2.resize(bm[0, 1].numpy(), (width, height)), (3, 3))
    ], axis=2)).unsqueeze(0)

    out = F.grid_sample(
        torch.from_numpy(image).permute(2, 0, 1).unsqueeze(0).float(),
        flow, align_corners=True
    )
    result = ((out[0] * 255).permute(1, 2, 0).numpy()[:, :, ::-1]).astype(np.uint8)

    out_dir = os.path.dirname(output_path)
    if out_dir:
        os.makedirs(out_dir, exist_ok=True)
    cv2.imwrite(output_path, result)
    print(f"[OK] {output_path}")


if __name__ == "__main__":
    main()