有参考图像的图像质量评估方法及代码（PSNR,SSIM,RMSE,NRMSE,ENTROPY）

Posted 2022-11-29 告白少年

1.峰值信噪比PSNR

• PSNR（Peak Signal to Noise Ratio），峰值信噪比，即峰值信号的能量与噪声的平均能量之比，通常表示的时候取
  log 变成分贝（dB），由于 MSE 为真实图像与含噪图像之差的能量均值，而两者的差即为噪声，因此 PSNR 即峰值信号能量与 MSE
  之比。

2.结构相似性SSIM

• 作为结构相似性理论的实现，结构相似度指数从图像组成的角度将结构信息定义为独立于亮度、对比度的，反映场景中物体结构的属性，并将失真建模为亮度、对比度和结构三个不同因素的组合。用均值作为亮度的估计，标准差作为对比度的估计，协方差作为结构相似程度的度量。
  
  
  

3.均方差RMSE

• 均方误差(Mean Square Error)反映的是变量间的差异程度，是一种基于像素误差的图像质量客观评价指标，用于衡量融合图像和理想参考图像之间的差异，MSE越小，表示融合图像质量越好。
  
• 均方根误差(Root Mean Square Error)是一个翻译空间细节信息的评价指标

4.归一化均方根误差 NRMSE

• 归一化均方根误差（normalized root mean square error）就是将RMSE的值变成(0,1)之间。

5.信息熵ENTROPY

import numpy as np
import math
import cv2
import os
from skimage.measure import compare_ssim
import skimage.measure
# 峰值信噪比
def psnr(target, ref):
	#将图像格式转为float64
    target_data = np.array(target, dtype=np.float64)
    ref_data = np.array(ref,dtype=np.float64)
    # 直接相减，求差值
    diff = ref_data - target_data
    # 按第三个通道顺序把三维矩阵拉平
    diff = diff.flatten('C')
    # 计算MSE值
    rmse = math.sqrt(np.mean(diff ** 2.))
    # 精度
    eps = np.finfo(np.float64).eps
    if(rmse == 0):
        rmse = eps
    psnr=20*math.log10(255.0/rmse)
    return psnr
# 结构相似性
def ssim(imageA, imageB):
    # 为确保图像能被转为灰度图
    imageA = np.array(imageA, dtype=np.uint8)
    imageB = np.array(imageB, dtype=np.uint8)

    # 通道分离，注意顺序BGR不是RGB
    (B1, G1, R1) = cv2.split(imageA)
    (B2, G2, R2) = cv2.split(imageB)

    # convert the images to grayscale BGR2GRAY
    grayA = cv2.cvtColor(imageA, cv2.COLOR_BGR2GRAY)
    grayB = cv2.cvtColor(imageB, cv2.COLOR_BGR2GRAY)

    # 方法一
    #直接转换为灰度图，计算ssim
    (grayScore, diff) = compare_ssim(grayA, grayB, full=True)
    diff = (diff * 255).astype("uint8")
    # print("gray SSIM: ".format(grayScore))

    # 方法二
    #分离每个通道，计算ssim，再求平均值
    (score0, diffB) = compare_ssim(B1, B2, full=True)
    (score1, diffG) = compare_ssim(G1, G2, full=True)
    (score2, diffR) = compare_ssim(R1, R2, full=True)
    aveScore = (score0 + score1 + score2) / 3
    # print("BGR average SSIM: ".format(aveScore))

    return  aveScore
#
##数据集评估####
original_path=r"C:\\study_he\\exp_result\\LOL15\\high"
contrast_path=r"C:\\study_he\\RetinexNet-master\\Retinex_n"
psnr_all=0
ssim_all=0
rmse_all=0
nrmse_all=0
entropy_all=0
length=len(os.listdir(original_path))
for file in os.listdir(original_path):
    original_img=os.path.join(original_path,file)
    contrast_img=os.path.join(contrast_path,file)
    original=cv2.imread(original_img)
    contrast = cv2.imread(contrast_img, 1)
    # 峰值信噪比
    psnrValue = psnr(original, contrast)
    # 结构相似性
    ssimValue = ssim(original, contrast)
    # 均方误差
    mse = skimage.measure.compare_mse(original, contrast)
    # 均方根误差
    rmse = math.sqrt(mse)
    # 归一化均方根误差
    nrmse = skimage.measure.compare_nrmse(original, contrast, norm_type='Euclidean')
    # 信息熵
    entropy = skimage.measure.shannon_entropy(contrast, base=2)
    psnr_all+=psnrValue
    ssim_all+=ssimValue
    rmse_all+=rmse
    nrmse_all+=nrmse
    entropy_all+=entropy

psnr_avg=psnr_all/length
ssim_avg=ssim_all/length
rmse_avg=rmse_all/length
nrmse_avg=nrmse_all/length
entropy_avg=entropy_all/length
print(rmse_avg)
print(nrmse_avg)
print(ssim_avg)
print(psnr_avg)
print(entropy_avg)

# #单张图像预测
# original = cv2.imread(r"C:\\study_he\\exp_result\\LOL15\\high\\780.png")      # numpy.adarray
# contrast = cv2.imread(r"C:\\study_he\\RetinexNet-master\\Retinex_n\\780.png",1)
# #峰值信噪比
# psnrValue = psnr(original,contrast)
# #结构相似性
# ssimValue = ssim(original,contrast)
# # 均方误差
# mse = skimage.measure.compare_mse(original, contrast)
# # 均方根误差
# rmse = math.sqrt(mse)
# # 归一化均方根误差
# nrmse = skimage.measure.compare_nrmse(original, contrast, norm_type='Euclidean')
# # 信息熵
# entropy = skimage.measure.shannon_entropy(contrast, base=2)
# print(rmse)
# print(nrmse)
# print(ssimValue)
# print(psnrValue)
# print(entropy)