超分辨入门之SRCNN（小白版）

《Learning a Deep Convolutional Network for Image Super-Resolution》

论文地址：Learning a Deep Convolutional Network for Image Super-Resolution | SpringerLink

        SRCNN的主要贡献是首次将深度学习引入超分辨任务中，改变了以往映射建模以及稀疏编码的超分辨方式，并且在测试集中证明了具有优越性能。

        在2023年的现在来开，SRCNN重建的效果已经远远不如使用残差学习、通道注意力机制或生成对抗网络的新结构，但是作为开山之作，这依旧是我这种小白作为接触超分辨的第一选择。

        好了，下面直接贴对应模块的代码以及相应数据集下载地址：

GitHub - yjn870/SRCNN-pytorch: PyTorch implementation of Image Super-Resolution Using Deep Convolutional Networks (ECCV 2014)https://github.com/yjn870/SRCNN-pytorch

一、models

from torch import nn


class SRCNN(nn.Module):
    def __init__(self, num_channels=1):
        super(SRCNN, self).__init__()
        self.conv1 = nn.Conv2d(num_channels, 64, kernel_size=9, padding=9 // 2)
        self.conv2 = nn.Conv2d(64, 32, kernel_size=5, padding=5 // 2)
        self.conv3 = nn.Conv2d(32, num_channels, kernel_size=5, padding=5 // 2)
        self.relu = nn.ReLU(inplace=True)

    def forward(self, x):
        x = self.relu(self.conv1(x))
        x = self.relu(self.conv2(x))
        x = self.conv3(x)
        return x

二、utils

import torch
import numpy as np


def convert_rgb_to_y(img):
    if type(img) == np.ndarray:
        return 16. + (64.738 * img[:, :, 0] + 129.057 * img[:, :, 1] + 25.064 * img[:, :, 2]) / 256.
    elif type(img) == torch.Tensor:
        if len(img.shape) == 4:
            img = img.squeeze(0)
        return 16. + (64.738 * img[0, :, :] + 129.057 * img[1, :, :] + 25.064 * img[2, :, :]) / 256.
    else:
        raise Exception('Unknown Type', type(img))


def convert_rgb_to_ycbcr(img):
    if type(img) == np.ndarray:
        y = 16. + (64.738 * img[:, :, 0] + 129.057 * img[:, :, 1] + 25.064 * img[:, :, 2]) / 256.
        cb = 128. + (-37.945 * img[:, :, 0] - 74.494 * img[:, :, 1] + 112.439 * img[:, :, 2]) / 256.
        cr = 128. + (112.439 * img[:, :, 0] - 94.154 * img[:, :, 1] - 18.285 * img[:, :, 2]) / 256.
        return np.array([y, cb, cr]).transpose([1, 2, 0])
    elif type(img) == torch.Tensor:
        if len(img.shape) == 4:
            img = img.squeeze(0)
        y = 16. + (64.738 * img[0, :, :] + 129.057 * img[1, :, :] + 25.064 * img[2, :, :]) / 256.
        cb = 128. + (-37.945 * img[0, :, :] - 74.494 * img[1, :, :] + 112.439 * img[2, :, :]) / 256.
        cr = 128. + (112.439 * img[0, :, :] - 94.154 * img[1, :, :] - 18.285 * img[2, :, :]) / 256.
        return torch.cat([y, cb, cr], 0).permute(1, 2, 0)
    else:
        raise Exception('Unknown Type', type(img))


def convert_ycbcr_to_rgb(img):
    if type(img) == np.ndarray:
        r = 298.082 * img[:, :, 0] / 256. + 408.583 * img[:, :, 2] / 256. - 222.921
        g = 298.082 * img[:, :, 0] / 256. - 100.291 * img[:, :, 1] / 256. - 208.120 * img[:, :, 2] / 256. + 135.576
        b = 298.082 * img[:, :, 0] / 256. + 516.412 * img[:, :, 1] / 256. - 276.836
        return np.array([r, g, b]).transpose([1, 2, 0])
    elif type(img) == torch.Tensor:
        if len(img.shape) == 4:
            img = img.squeeze(0)
        r = 298.082 * img[0, :, :] / 256. + 408.583 * img[2, :, :] / 256. - 222.921
        g = 298.082 * img[0, :, :] / 256. - 100.291 * img[1, :, :] / 256. - 208.120 * img[2, :, :] / 256. + 135.576
        b = 298.082 * img[0, :, :] / 256. + 516.412 * img[1, :, :] / 256. - 276.836
        return torch.cat([r, g, b], 0).permute(1, 2, 0)
    else:
        raise Exception('Unknown Type', type(img))


def calc_psnr(img1, img2):
    return 10. * torch.log10(1. / torch.mean((img1 - img2) ** 2))


class AverageMeter(object):
    def __init__(self):
        self.reset()

    def reset(self):
        self.val = 0
        self.avg = 0
        self.sum = 0
        self.count = 0

    def update(self, val, n=1):
        self.val = val
        self.sum += val * n
        self.count += n
        self.avg = self.sum / self.count

三、prepare

import argparse
import glob
import h5py
import numpy as np
import PIL.Image as pil_image
from utils import convert_rgb_to_y

import sys
sys.argv=['']
del sys

def train(args):
    h5_file = h5py.File(args.output_path, 'w')

    lr_patches = []
    hr_patches = []

    for image_path in sorted(glob.glob('{}/*'.format(args.images_dir))):
        hr = pil_image.open(image_path).convert('RGB')
        hr_width = (hr.width // args.scale) * args.scale
        hr_height = (hr.height // args.scale) * args.scale
        hr = hr.resize((hr_width, hr_height), resample=pil_image.BICUBIC)
        lr = hr.resize((hr_width // args.scale, hr_height // args.scale), resample=pil_image.BICUBIC)
        lr = lr.resize((lr.width * args.scale, lr.height * args.scale), resample=pil_image.BICUBIC)
        hr = np.array(hr).astype(np.float32)
        lr = np.array(lr).astype(np.float32)
        hr = convert_rgb_to_y(hr)
        lr = convert_rgb_to_y(lr)

        for i in range(0, lr.shape[0] - args.patch_size + 1, args.stride):
            for j in range(0, lr.shape[1] - args.patch_size + 1, args.stride):
                lr_patches.append(lr[i:i + args.patch_size, j:j + args.patch_size])
                hr_patches.append(hr[i:i + args.patch_size, j:j + args.patch_size])

    lr_patches = np.array(lr_patches)
    hr_patches = np.array(hr_patches)

    h5_file.create_dataset('lr', data=lr_patches)
    h5_file.create_dataset('hr', data=hr_patches)

    h5_file.close()


def eval(args):
    h5_file = h5py.File(args.output_path, 'w')

    lr_group = h5_file.create_group('lr')
    hr_group = h5_file.create_group('hr')

    for i, image_path in enumerate(sorted(glob.glob('{}/*'.format(args.images_dir)))):
        hr = pil_image.open(image_path).convert('RGB')
        hr_width = (hr.width // args.scale) * args.scale
        hr_height = (hr.height // args.scale) * args.scale
        hr = hr.resize((hr_width, hr_height), resample=pil_image.BICUBIC)
        lr = hr.resize((hr_width // args.scale, hr_height // args.scale), resample=pil_image.BICUBIC)
        lr = lr.resize((lr.width * args.scale, lr.height * args.scale), resample=pil_image.BICUBIC)
        hr = np.array(hr).astype(np.float32)
        lr = np.array(lr).astype(np.float32)
        hr = convert_rgb_to_y(hr)
        lr = convert_rgb_to_y(lr)

        lr_group.create_dataset(str(i), data=lr)
        hr_group.create_dataset(str(i), data=hr)

    h5_file.close()


if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('--images-dir', type=str, default='D:\MyLearning\SRCNN-pytorch-master\DataSet')
    parser.add_argument('--output-path', type=str, default='D:\MyLearning\SRCNN-pytorch-master\outPut')
    parser.add_argument('--patch-size', type=int, default=33)
    parser.add_argument('--stride', type=int, default=14)
    parser.add_argument('--scale', type=int, default=2)
    parser.add_argument('--eval', action='store_true')
    parser.add_argument('-f', type=str, default="读取额外的参数")
    args = parser.parse_args(args=[])

    if not args.eval:
        train(args)
    else:
        eval(args)

四、dataset

import h5py
import numpy as np
from torch.utils.data import Dataset


class TrainDataset(Dataset):
    def __init__(self, h5_file):
        super(TrainDataset, self).__init__()
        self.h5_file = h5_file

    def __getitem__(self, idx):
        with h5py.File(self.h5_file, 'r') as f:
            return np.expand_dims(f['lr'][idx] / 255., 0), np.expand_dims(f['hr'][idx] / 255., 0)

    def __len__(self):
        with h5py.File(self.h5_file, 'r') as f:
            return len(f['lr'])


class EvalDataset(Dataset):
    def __init__(self, h5_file):
        super(EvalDataset, self).__init__()
        self.h5_file = h5_file

    def __getitem__(self, idx):
        with h5py.File(self.h5_file, 'r') as f:
            return np.expand_dims(f['lr'][str(idx)][:, :] / 255., 0), np.expand_dims(f['hr'][str(idx)][:, :] / 255., 0)

    def __len__(self):
        with h5py.File(self.h5_file, 'r') as f:
            return len(f['lr'])

五、train

需要注意，这里scale的值需要与你下载的训练集保持一致

import argparse
import os
import copy

import torch
from torch import nn
import torch.optim as optim
import torch.backends.cudnn as cudnn
from torch.utils.data.dataloader import DataLoader
from tqdm import tqdm

from models import SRCNN
from datasets import TrainDataset, EvalDataset
from utils import AverageMeter, calc_psnr


if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    #替换为本地路径
    parser.add_argument('--train-file', type=str, default='D:\MyLearning\SRCNN-pytorch-master\91-image_x3.h5')
    parser.add_argument('--eval-file', type=str, default='D:\MyLearning\SRCNN-pytorch-master\Set5_x3.h5')
    parser.add_argument('--outputs-dir', type=str, default='D:\MyLearning\SRCNN-pytorch-master\output')
    parser.add_argument('--scale', type=int, default=3)
    parser.add_argument('--lr', type=float, default=1e-4)
    parser.add_argument('--batch-size', type=int, default=16)
    parser.add_argument('--num-epochs', type=int, default=400)
    parser.add_argument('--num-workers', type=int, default=0)
    parser.add_argument('--seed', type=int, default=123)
    args = parser.parse_args()

    args.outputs_dir = os.path.join(args.outputs_dir, 'x{}'.format(args.scale))

    if not os.path.exists(args.outputs_dir):
        os.makedirs(args.outputs_dir)

    cudnn.benchmark = True
    device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')

    torch.manual_seed(args.seed)

    model = SRCNN().to(device)
    criterion = nn.MSELoss()
    optimizer = optim.Adam([
        {'params': model.conv1.parameters()},
        {'params': model.conv2.parameters()},
        {'params': model.conv3.parameters(), 'lr': args.lr * 0.1}
    ], lr=args.lr)

    train_dataset = TrainDataset(args.train_file)
    train_dataloader = DataLoader(dataset=train_dataset,
                                  batch_size=args.batch_size,
                                  shuffle=True,
                                  num_workers=args.num_workers,
                                  pin_memory=True,
                                  drop_last=True)
    eval_dataset = EvalDataset(args.eval_file)
    eval_dataloader = DataLoader(dataset=eval_dataset, batch_size=1)

    best_weights = copy.deepcopy(model.state_dict())
    best_epoch = 0
    best_psnr = 0.0

    for epoch in range(args.num_epochs):
        model.train()
        epoch_losses = AverageMeter()

        with tqdm(total=(len(train_dataset) - len(train_dataset) % args.batch_size)) as t:
            t.set_description('epoch: {}/{}'.format(epoch, args.num_epochs - 1))

            for data in train_dataloader:
                inputs, labels = data

                inputs = inputs.to(device)
                labels = labels.to(device)

                preds = model(inputs)

                loss = criterion(preds, labels)

                epoch_losses.update(loss.item(), len(inputs))

                optimizer.zero_grad()
                loss.backward()
                optimizer.step()

                t.set_postfix(loss='{:.6f}'.format(epoch_losses.avg))
                t.update(len(inputs))

        torch.save(model.state_dict(), os.path.join(args.outputs_dir, 'epoch_{}.pth'.format(epoch)))

        model.eval()
        epoch_psnr = AverageMeter()

        for data in eval_dataloader:
            inputs, labels = data

            inputs = inputs.to(device)
            labels = labels.to(device)

            with torch.no_grad():
                preds = model(inputs).clamp(0.0, 1.0)

            epoch_psnr.update(calc_psnr(preds, labels), len(inputs))

        print('eval psnr: {:.2f}'.format(epoch_psnr.avg))

        if epoch_psnr.avg > best_psnr:
            best_epoch = epoch
            best_psnr = epoch_psnr.avg
            best_weights = copy.deepcopy(model.state_dict())

    print('best epoch: {}, psnr: {:.2f}'.format(best_epoch, best_psnr))
    torch.save(best_weights, os.path.join(args.outputs_dir, 'best.pth'))

六、test

import argparse

import torch
import torch.backends.cudnn as cudnn
import numpy as np
import PIL.Image as pil_image

from models import SRCNN
from utils import convert_rgb_to_ycbcr, convert_ycbcr_to_rgb, calc_psnr


if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    #路径需要替换为自己训练生成的权重文件
    parser.add_argument('--weights-file', type=str, default='outPut\\x3\\best.pth')
    #路径需要替换为需要进行测试的图片路径
    parser.add_argument('--image-file', type=str,default='D:\\MyLearning\\SRCNN-pytorch-master\\5_1_19.bmp')
    parser.add_argument('--scale', type=int, default=3)
    args = parser.parse_args()

    cudnn.benchmark = True
    device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')

    model = SRCNN().to(device)

    state_dict = model.state_dict()
    for n, p in torch.load(args.weights_file, map_location=lambda storage, loc: storage).items():
        if n in state_dict.keys():
            state_dict[n].copy_(p)
        else:
            raise KeyError(n)

    model.eval()
    image = pil_image.open(args.image_file).convert('RGB')
    
    image_width = (image.width // args.scale) * args.scale
    image_height = (image.height // args.scale) * args.scale
    image = image.resize((image_width, image_height), resample=pil_image.BICUBIC)
    image = image.resize((image.width // args.scale, image.height // args.scale), resample=pil_image.BICUBIC)
    image = image.resize((image.width * args.scale, image.height * args.scale), resample=pil_image.BICUBIC)
    image.save(args.image_file.replace('.', '_bicubic_x{}.'.format(args.scale)))

    image = np.array(image).astype(np.float32)
    ycbcr = convert_rgb_to_ycbcr(image)

    y = ycbcr[..., 0]
    y /= 255.
    y = torch.from_numpy(y).to(device)
    y = y.unsqueeze(0).unsqueeze(0)

    with torch.no_grad():
        preds = model(y).clamp(0.0, 1.0)

    psnr = calc_psnr(y, preds)
    print('PSNR: {:.2f}'.format(psnr))

    preds = preds.mul(255.0).cpu().numpy().squeeze(0).squeeze(0)

    output = np.array([preds, ycbcr[..., 1], ycbcr[..., 2]]).transpose([1, 2, 0])
    output = np.clip(convert_ycbcr_to_rgb(output), 0.0, 255.0).astype(np.uint8)
    output = pil_image.fromarray(output)
    output.save(args.image_file.replace('.', '_srcnn_x{}.'.format(args.scale)))