快速高速模糊算法

刚才发现一份快速高速模糊的实现。

源地址为：http://incubator.quasimondo.com/processing/gaussian_blur_1.php

作者信息为: Fast Gaussian Blur v1.3
 by Mario Klingemann <http://incubator.quasimondo.com>

processing源码: http://incubator.quasimondo.com/processing/fastblur.pde效果图：

转为C语言实现版本。

代码如下:

// Fast Gaussian Blur v1.3
// by Mario Klingemann <http://incubator.quasimondo.com>
// C version updated and performance optimization by tntmonks(http://tntmonks.cnblogs.com)
// One of my first steps with Processing. I am a fan
// of blurring. Especially as you can use blurred images
// as a base for other effects. So this is something I
// might get back to in later experiments.
//
// What you see is an attempt to implement a Gaussian Blur algorithm
// which is exact but fast. I think that this one should be
// relatively fast because it uses a special trick by first
// making a horizontal blur on the original image and afterwards
// making a vertical blur on the pre-processed image. This
// is a mathematical correct thing to do and reduces the
// calculation a lot.
//
// In order to avoid the overhead of function calls I unrolled
// the whole convolution routine in one method. This may not
// look nice, but brings a huge performance boost.
//
//
// v1.1: I replaced some multiplications by additions
//       and added aome minor pre-caclulations.
//       Also add correct rounding for float->int conversion
//
// v1.2: I completely got rid of all floating point calculations
//       and speeded up the whole process by using a
//       precalculated multiplication table. Unfortunately
//       a precalculated division table was becoming too
//       huge. But maybe there is some way to even speed
//       up the divisions.
//
// v1.3: Fixed a bug that caused blurs that start at y>0
//	 to go wrong. Thanks to Jeroen Schellekens for
//       finding it!

//申请二维数组
template <typename T>
T** newArray2D(unsigned int width, unsigned int height)
{
	unsigned int size = sizeof(T);
	unsigned int point_size = sizeof(T*);

	T **arr = (T **)malloc(point_size * width + size * width * height);
	memset(arr, 0, point_size * width + size * width * height);
	if (arr != NULL)
	{
		T *head = (T*)((int)arr + point_size * width);
		for (unsigned int i = 0; i < width; ++i)
		{
			arr[i] = (T*)((int)head + i * height * size);
			for (unsigned int j = 0; j < height; ++j)
				new (&arr[i][j]) T;
		}
	}
	return (T**)arr;
}
//释放二维数组
template <typename T>
void deleteArray2D(T **arr, unsigned int width, unsigned int height)
{
	for (unsigned int i = 0; i < width; ++i)
		for (unsigned int j = 0; j < height; ++j)
			arr[i][j].~T();
	if (arr != NULL)
		free((void**)arr);
} 

void GaussianBlur(unsigned char* img, unsigned  int x, unsigned int y, unsigned int w, unsigned int h,unsigned int comp, unsigned int radius)
{
	unsigned int i, j;
	radius = min(max(1, radius), 248);

	unsigned int kernelSize = 1 + radius * 2;

	unsigned int* kernel = (unsigned int*)malloc(kernelSize* sizeof(unsigned int));
	memset(kernel, 0, kernelSize* sizeof(unsigned int));
	unsigned int** mult = newArray2D<unsigned int>(kernelSize, 256);

	unsigned	int sum = 0;
	for (i = 1; i < radius; i++){
		unsigned int szi = radius - i;
		kernel[radius + i] = kernel[szi] = szi*szi;
		sum += kernel[szi] + kernel[szi];
		for (j = 0; j < 256; j++){
			mult[radius + i][j] = mult[szi][j] = kernel[szi] * j;
		}
	}
	kernel[radius] = radius*radius;
	sum += kernel[radius];
	for (j = 0; j < 256; j++){
		mult[radius][j] = kernel[radius] * j;
	}

	unsigned int   cr, cg, cb;
	unsigned int   xl, yl, yi, ym, riw;
	int   read, ri;
	unsigned	int imgWidth =  w;
	unsigned	int imgHeight =  h;
	unsigned	int imageSize = imgWidth*imgHeight;
	unsigned char *	rgb = (unsigned char *)malloc(sizeof(unsigned char) * imageSize*3);
	unsigned char *	r = rgb;
	unsigned char *	g = rgb + imageSize;
	unsigned char *	b = rgb + imageSize*2;
	unsigned char *	rgb2 = (unsigned char *)malloc(sizeof(unsigned char) * imageSize * 3);
	unsigned char *	r2 = rgb2;
	unsigned char *	g2 = rgb2 + imageSize;
	unsigned char *	b2 = rgb2 + imageSize * 2;
	for (size_t yh = 0; yh < imgHeight; ++yh) {
		for (size_t xw = 0; xw < imgWidth; ++xw) {
			unsigned int n = xw + yh* imgWidth;
			unsigned int p = n*comp;
			r[n] = img[p];
			g[n] = img[p + 1];
			b[n] = img[p + 2];
		}
	}
	unsigned int Zero = 0;
	//max(0, x);
	x = x ^ ((x ^ Zero) & -(x < Zero));
	//max(0, y);
	y = Zero ^ ((Zero ^ y) & -(Zero < y)); 

	w = x + w - max(0, (x + w) - imgWidth);
	h = y + h - max(0, (y + h) - imgHeight);
	yi = y*imgWidth;

	for (yl = y; yl < h; yl++){
		for (xl = x; xl < w; xl++){
			cb = cg = cr = sum = 0;
			ri = xl - radius;
			for (i = 0; i < kernelSize; i++){
				read = ri + i;
			 	if (read >= x && read < w)
				{
					read += yi;
					cr += mult[i][r[read]];
					cg += mult[i][g[read]];
					cb += mult[i][b[read]];
					sum += kernel[i];
				}
			}
			ri = yi + xl;
			r2[ri] = cr / sum;
			g2[ri] = cg / sum;
			b2[ri] = cb / sum;
		}
		yi += imgWidth;
	}
	yi = y*imgWidth;

	for (yl = y; yl < h; yl++){
		ym = yl - radius;
		riw = ym*imgWidth;
		for (xl = x; xl < w; xl++){
			cb = cg = cr = sum = 0;
			ri = ym;
			read = xl + riw;
			for (i = 0; i < kernelSize; i++){
				 	if (ri < h && ri >= y)
				{
					cr += mult[i][r2[read]];
					cg += mult[i][g2[read]];
					cb += mult[i][b2[read]];
					sum += kernel[i];
				}
				ri++;
				read += imgWidth;
			}
			unsigned int p = (xl + yi)*comp;
			img[p] = (unsigned char)(cr / sum);
			img[p + 1] = (unsigned char)(cg / sum);
			img[p + 2] = (unsigned char)(cb / sum);
		}
		yi += imgWidth;
	}

	free(rgb);
	free(rgb2);
	free(kernel);
	deleteArray2D(mult, 1 + radius * 2, 256);
}

　　该代码，将二维数组进一步优化后可提升一定的速度。

在博主机子上测试一张5000x3000的图像，模糊半径为10的情况下，耗时4s.