数据压缩实验四——DPCM压缩系统的实现和分析

DPCM压缩系统的实现和分析

---

一、实验原理

DPCM是差分预测编码调制的缩写，本次实验采用固定预测器和均匀量化器，预测器采用前向预测。其中xn为输入信号，Q为量化器，dn=pn+xn为预测误差，P为延时器。

大致算法步骤如下：
读取一个256级的灰度图像，与前一个像素作差得到预测误差，范围为[-255,255]。
需要将其范围转变为计算机可以表示的范围，先转变为单极性，即加上255，范围变为[0,510]。
进行8bit均匀量化。

二、实验步骤

1.主函数读写文件

int main(void)
{
	int bits_depth = 8;
	unsigned char* yBuf, * uBuf, * vBuf;
	int w = 256, h = 256;
	yBuf = (unsigned char*)malloc(sizeof(unsigned char) * (w * h));
	uBuf = (unsigned char*)malloc(sizeof(unsigned char) * (w * h / 4));
	vBuf = (unsigned char*)malloc(sizeof(unsigned char) * (w * h / 4));
	FILE* src = NULL;
	src = fopen("Lena256B.yuv", "rb");
	if (src == NULL)
	{
		printf("error opening source file!\n");
		system("pause");
		exit(-1);
	}
	fread(yBuf, sizeof(unsigned char), w * h, src);
	fread(uBuf, sizeof(unsigned char), w * h / 4, src);
	fread(vBuf, sizeof(unsigned char), w * h / 4, src);
	fclose(src);
	unsigned char* qBuf, * reBuf;
	qBuf = (unsigned char*)malloc(sizeof(unsigned char) * (w * h));//预测误差buffer
	reBuf = (unsigned char*)malloc(sizeof(unsigned char) * (w * h));//重构图像buffer


	dpcm(yBuf, qBuf, reBuf, w, h, bits_depth);


	FILE* obj1 = NULL;
	FILE* obj2 = NULL;
	obj1 = fopen("q_Lena256B.yuv", "wb");
	obj2 = fopen("re_Lena256B.yuv", "wb");
	if (obj1 == NULL || obj2 == NULL)
	{
		printf("error opening objective file!\n");
		system("pause");
		exit(-1);
	}
	fwrite(qBuf, sizeof(unsigned char), w * h, obj1);
	fwrite(uBuf, sizeof(unsigned char), w * h / 4, obj1);
	fwrite(vBuf, sizeof(unsigned char), w * h / 4, obj1);
	fwrite(reBuf, sizeof(unsigned char), w * h, obj2);
	fwrite(uBuf, sizeof(unsigned char), w * h / 4, obj2);
	fwrite(vBuf, sizeof(unsigned char), w * h / 4, obj2);
	fclose(obj1);
	fclose(obj2);

2.进行DPCM

void dpcm(unsigned char* yBuf, unsigned char* qBuf, unsigned char* reBuf, int w, int h, int depth)
{
	int r = pow(2, (double)(9 - depth));
	for (int i = 0; i < h; i++)
		for (int j = 0; j < w; j++)
		{
			if (j == 0)//第一列保持原样值不变
			{
				qBuf[i * w] = ((yBuf[i * w] - 128) + 255) / pow(2, (double)(9 - depth));//预测误差量化
				reBuf[i * w] = qBuf[i * w] * pow(2, (double)(9 - depth)) + 255 - 128;//反量化
				
			}
			else
			{
				qBuf[i * w + j] = ((yBuf[i * w + j] - reBuf[i * w + j - 1]) + 255) / pow(2, (double)(9 - depth));//当前预测误差为当前样值与上个重构样值的差值（量化后）
				reBuf[i * w + j] = qBuf[i * w + j] * pow(2, (double)(9 - depth)) - 255 + reBuf[i * w + j - 1];//当前重构样值为量化后的当前预测误差与上一个重构样值的和

			}
			if (qBuf[i * w + j] > 255)
				qBuf[i * w + j] = 255;
			if (qBuf[i * w + j] < 0)
				qBuf[i * w + j] = 0;
			if (reBuf[i * w + j] > 255)
				reBuf[i * w + j] = 255;
			if (reBuf[i * w + j] < 0)
				reBuf[i * w + j] = 0;
		}
}

3.计算概率分布

//计算分布概率
	double fre_q[256] = { 0 };
	double fre_org[256] = { 0 };
	for (int i = 0; i < 256; i++)
	{
		int count_q = 0, count_org = 0;
		for (int j = 0; j < w * h; j++)
		{
			if ((int)qBuf[j] == i)
				count_q++;
			if ((int)yBuf[j] == i)
				count_org++;
		}
		fre_q[i] = (double)count_q / w / h;
		fre_org[i] = (double)count_org / w / h;
	}
	char s[] = "symbol\tfrequency\n";
	FILE* data_q = fopen("q_Lena256B.txt", "w");
	FILE* data_org = fopen("org_Lena256B.txt", "w");
	if (data_q == NULL || data_org == NULL)
	{
		printf("error opening txt file!\n");
		system("pause");
		exit(-1);
	}
	fprintf(data_org, s);
	fprintf(data_q, s);
	for (int i = 0;i < 256; i++)
	{
		fprintf(data_q, "%d\t%f\n", i, fre_q[i]);
		fprintf(data_org, "%d\t%f\n", i, fre_org[i]);
	}
	fclose(data_q);
	fclose(data_org);

3.计算PSNR

//计算PSNR
	int max = 255;
	double mse = 0;
	for (int i = 0; i < h; i++)
		for (int j = 0; j < w; j++)
		{
			mse += (yBuf[i * w + j] - reBuf[i * w + j]) * (yBuf[i * w + j] - reBuf[i * w + j]);
		}
	mse = (double)mse / (double)(w * h);
	double psnr = 10 * log10((double)(max * max) / mse);
	cout << "PSNR = " << psnr;
	system("pause");

三、实验结果

1.图像

8bit 量化时，原图像、预测误差图像、重建图像如下：
原图像、2bit、4bit、8bit量化的重建图像

2.PSNR结果

PSNR	2bit	4bit	8bit
Lena.yuv	7.66699	14.6621	50.9301

PSNR随量化bit数增加而增大，重建图像质量也越好。

3.概率分布曲线

原图：
量化误差图像概率分布曲线：
8bit
4bit
2bit

4.用huffcode对量化误差图像进行压缩

使用Huffcode.exe处理量化误差图像后，图像被压缩。

图像	压缩前（yuv）kb	压缩后（huff）kb	压缩比
Lena256B.yuv	96	69	71.86%
Lena256B.yuv(8bit)	96	46	47.92%
Lena256B.yuv(4bit)	96	24	25.00%
Lena256B.yuv(2bit)	96	23	23.96%

可见进行DPCM编码后的图像再进行熵编码，比仅熵编码的图像数据量小，且量化比特数越大，压缩比越小。

四、实验总结

1bit 量化时量化误差图像是黑的，量化比特数越低图像质量越差，DPCM编码可以有效地压缩文件。