I want to test on Levin, However, the results seem to be failed.

[LiuzhuForFun]

Can you give me some suggestions? `

for im_index in range(4): gt_name = r'data/levindata/img%d.png' % (im_index + 1) gt = cv2.imread(gt_name,0) gt = gt / 255

gt= gty[..., np.newaxis]

    # gt = np.concatenate([gt,gt,gt],axis=2)
    for ker_index in range(8):
        im_name = r'data/levin_data/im_%d_ker_%d.png' % (im_index + 1, ker_index + 1)
        ker_name = r'data/levin_data/kernel_%d.dlm' % (ker_index + 1)
        ker = np.loadtxt(ker_name).astype(np.float)
        ker = np.clip(ker, 0, 1)
        ker = ker / np.sum(ker)
        y =cv2.imread(im_name)/255.0
        # y = ndimage.convolve(gt, ker, mode='wrap')
        np.random.seed(1512818945)
        y = y
        y = np.clip(y, 0, 1.0)
        y = ((y * 255.0).astype(np.uint8) / 255.0).astype(np.float32)

        # y = np.concatenate([y,y,y],axis=2)
        # y = y+ test_sigma * np.random.standard_normal(size=y.shape)
        y = to_tensor(y,device)
        x = to_tensor(gt,device)
        k = to_tensor(ker,device)
        kt =torch.flip(k, [1, 2])
        y, kt, k = y.to(device), kt.to(device), k.to(device)

        k_size = k.size()[2]
        padding_size = int((k_size / 2) * 1.5)
        y = conv2d_same_padding(x, k).to(device)
        # nl = np.random.uniform(0.003, 0.015)
        y = y + 0.01 * torch.randn_like(y).to(device)
        y = torch.clamp(y * 255.0, 0, 255)
        y = y.type(torch.ByteTensor)
        y = y.float() / 255.0
        y = y.to(device)

        y = torch.nn.functional.pad(y, [padding_size, padding_size, padding_size, padding_size], mode='replicate')

        start = time.time()
        y = torch.cat([y,y,y],dim=1)
        output_seq = trained_model(y, k, kt)
        # output_len = len(output_seq)
        x_final = output_seq[-1]
        # print('Time {}'.format(time.time() - start))

        ##
        # if (opt.ImgPad):
        y = utils.truncate_image(y, padding_size)
        x_final = utils.truncate_image(x_final, padding_size)

        x_est_np = x_final.cpu()

        print(np.shape(x_est_np))
        x_est_np = utils.tensor_to_np_img(x_est_np)
        #

        x_est_np = utils.box_proj(x_est_np)

        # sample_name_full = sample_name[0]
        # sample_name = sample_name_full[0:len(sample_name_full) - 4]
        x_est_np = np.array(x_est_np).astype(np.float)
        PSNR_list.append(psnr(gt,x_est_np[:,:,0]))
        SSIM_list.append(compute_ssim(gt,x_est_np[:,:,0],L=1.0))
        # sio.savemat(outpath + sample_name + '_estx.mat', {'x_est': x_est_np})
        scipy.misc.imsave('./data/levin_test/' + str(im_index+1)+'_'+str(ker_index+1) + '_estx.png', x_est_np * 255)
        torch.cuda.empty_cache()

`

[LiuzhuForFun]

@donggong1 Thanks

[LiuzhuForFun]

On the gray scale of levin, the test results are 10.92dB and 0.085 SSIM