一、PSNR:峰值信噪比,公式如下:
其中MSE是原图像和处理图像的均方误差,公式如下:
PSNR越大,表示两图相似度越高
实现代码如下:
import torch import torch.nn.functional as F from math import log10 import cv2 import numpy as np import torchvision from skimage.metrics import structural_similarity as ssim def to_psnr(frame_out, gt): mse = F.mse_loss(frame_out, gt, reduction='none') mse_split = torch.split(mse, 1, dim=0) mse_list = [torch.mean(torch.squeeze(mse_split[ind])).item() for ind in range(len(mse_split))] intensity_max = 1.0 psnr_list = [10.0 * log10(intensity_max / mse) for mse in mse_list] return psnr_list
二、SSIM是一种结构性损失,通过衡量考察图与参考图在每个patch上的均值、标准差、协方差(以patch中心为中心的高斯窗加权求统计值)来衡量两张图的结构相似性。
skimage 代码实现:
def to_ssim_skimage(dehaze, gt): dehaze_list = torch.split(dehaze, 1, dim=0) gt_list = torch.split(gt, 1, dim=0) dehaze_list_np = [dehaze_list[ind].permute(0, 2, 3, 1).data.cpu().numpy().squeeze() for ind in range(len(dehaze_list))] gt_list_np = [gt_list[ind].permute(0, 2, 3, 1).data.cpu().numpy().squeeze() for ind in range(len(dehaze_list))] ssim_list = [ssim(dehaze_list_np[ind], gt_list_np[ind], data_range=1, multichannel=True) for ind in range(len(dehaze_list))] return ssim_list