图像傅里叶变换

对图像进行傅里叶变换，分出低频和高频部分。

image.png

def compute_fft(args, dataloader):
    #device = args.device if args.device == "cpu" else int(args.device)
    device = 0  # Do not modify
    h, w = args.img_size, args.img_size

    lpf = torch.zeros((h, w))
    R = (h + w) // 8
    for x in range(w):
        for y in range(h):
            if ((x - w/2)**2 + (y - h/2)**2) < R**2:
                lpf[y, x] = 1
    hpf = 1 - lpf
    hpf, lpf = hpf.to(device), lpf.to(device)

    with torch.no_grad():
        for i, (image_name, img, lbl) in enumerate(dataloader):
            print("img: %d" % i, image_name, 'label:', lbl)

            img = img.to(device)
            lbl = lbl.to(device)
            # print('img size', img.size())  # [1, 3, 224, 224]
            # img = normalize(img)

            fft_img = torch.fft.fftn(img, dim=(2, 3))
            # print('fft size', fft_img.size()) # [1, 3, 224, 224]
            # put low_freq into the center of img
            fft_img = torch.roll(fft_img, (h//2, w//2), dims=(2, 3))
            f_low = fft_img * lpf
            f_high = fft_img * hpf
            X_low = torch.abs(torch.fft.ifftn(f_low, dim=(2,3)))
            X_high = torch.abs(torch.fft.ifftn(f_high, dim=(2,3)))
            X_low_pil = transforms.ToPILImage()((torch.squeeze(X_low, dim=0)).float())
            X_low_pil.save('low_' + image_name[0])
            X_high_pil = transforms.ToPILImage()((torch.squeeze(X_high, dim=0)).float())
            X_high_pil.save('high_' + image_name[0])
           

着重参考->分别使用numpy和pytorch进行图像傅里叶变换和频域分析
图像处理之高通滤波及低通滤波_ReWz的博客-CSDN博客