视频质量评价PSNR
一幅MxN尺寸的图像的PSNR的计算公式如下所示：
其中xij 和yij 分别表示失真图像和原始图像对应像素点的灰度值；
i，j 分别代表图像的行和列；
L 是图像灰度值可到达的动态范围，8位的灰度图像的L=2^8-1=255。
如果已知SSD，MxN尺寸图像的PSNR公式如下所示。
PSNR=10*lg(255^2/MSE)
例如下图两张1080图片（左边是原图，右边是编码之后的图片，QP为17）的PSNR对比的结果 Y PSNR is 40.632022, U PSNR is 44.596545,V PSNR is 45.759277。
PSNR对比测试：
1080P视频测试：BasketballDrive_1920x1080_25_250.yuv
dp:17
Y PSNR is 40.632022, U PSNR is 44.596545,V PSNR is 45.759277
dp:28
Y PSNR is 38.834869, U PSNR is 42.445172,V PSNR is 42.977148
dp:32
Y PSNR is 38.144906, U PSNR is 41.265455,V PSNR is 41.705576
dp:36
Y PSNR is 37.527187, U PSNR is 40.110563,V PSNR is 40.551498
我们知道量化和反量化过程中，量化步长QP决定量化器的编码压缩率及图像精度。如果QP比较大，则量化值FQ动态范围越小，其相应的编码长度越小，但反量化是损失较多的图像细节，导致PSNR值越小。
代码：
// PSNR_CAL.cpp : 定义控制台应用程序的入口点。
//PSNR (Peak Signal to Noise Ratio)
//峰值信噪比PSNR衡量图像失真或是噪声水平的客观标准。2个图像之间PSNR值越大，则越相似。普遍基准为30dB，30dB以下的图像劣化较为明显。
#include "stdafx.h"
#include
#include
#include
#include
#define VIDEO_WIDTH 1280
#define VIDEO_HEIGHT 720
#define VIDEO_FRAME_NUM 10 //frame number
//编码解码YUV
#define DEC_YUV_PATH "E:\\yuv\\\\bs3.yuv"
//原YUV
#define REF_YUV_PATH "E:\\yuv\\BasketballDrive_1920x1080_25_250.yuv"
#define REF_YUV_422 0 //1:reference yuv is 422 format, 0:reference yuv is 420 format
#define VIDEO_SIZE_Y VIDEO_HEIGHT*VIDEO_WIDTH
#define VIDEO_SIZE_UV (VIDEO_HEIGHT*VIDEO_WIDTH)>>1
#define VIDEO_SIZE_YUV (VIDEO_SIZE_Y + VIDEO_SIZE_UV)
#define CONV422 0
#define CAL_PSNR 1
int main()
{
FILE *fp_dec;
FILE *fp_ref;
int i, j, k, comp;
#if CAL_PSNR
unsigned char line_dec[5000];
unsigned char line_ref[5000];
int idiff;
unsigned long diff_sum;
int width, height;
double psnr_frame;
double psnr_sum[VIDEO_FRAME_NUM][3];
double psnr_total[3];
fp_dec = fopen(DEC_YUV_PATH, "rb");
fp_ref = fopen(REF_YUV_PATH, "rb");
if (fp_dec == NULL)
{
printf("\n DEC YUV file not found\n");
return 0;
}
if (fp_ref == NULL)
{
printf("\n REF YUV file not found\n");
return 0;
}
for (i = 0; i < VIDEO_FRAME_NUM; i++)
{ //Y
for (comp = 0; comp < 3; comp++)
{
diff_sum = 0;
if(comp ==0)
{
width = VIDEO_WIDTH;
height = VIDEO_HEIGHT;
}
else
{
width = VIDEO_WIDTH /2;
height = VIDEO_HEIGHT /2;
}
for (j = 0; j < height; j++)
{
fread(line_dec, 1, width, fp_dec);
fread(line_ref, 1, width, fp_ref);
//fwrite(line_ref, 1, width, fp_ref_422);
//if(comp != 0) //UV
// fwrite(line_ref, 1, width, fp_ref_422);
for (k = 0; k < width; k++)
{
idiff = (int)(line_dec[k] - line_ref[k]);
diff_sum += idiff*idiff;
//if (k == 0 && j == 5 )
// printf("stop at %d", k);
}
//if (comp != 0 && REF_YUV_422 == 1) // if 422 format, skip one chroma line
// fread(line_dec, 1, width, fp_dec);
}
psnr_frame = (double)255 * 255 * width* height;
psnr_sum[i][comp] = 10.0 * log10(psnr_frame / (double)diff_sum);
}
printf("frame %d, Y PSNR is %f, Cb PSNR is %f,Cr PSNR is %f \n", i, psnr_sum[i][0], psnr_sum[i][1], psnr_sum[i][2]);
}
psnr_total[0] = 0;
psnr_total[1] = 0;
psnr_total[2] = 0;
for (i = 0; i 但是PSNR仅仅计算了图像像素点间的绝对误差，没有考虑像素点间的视觉相关性，更没顾及人类视觉系统的感知特性，所以其评价结果与主观感受往往相差较大(SSIM就是一种典型的与人类视觉系统特性结合的质量评价方法)。