Double-sided filter

Question

How to implement a two-way filter, given the Gaussian filter?

Answer 1

A simple two-way filter can be defined as

Inew (x, y) = Summation (j = yn / 2; j <= y + n / 2) Summation (i = xm / 2; j <= x + m / 2) w (i, j, x, y ) I (I, J)

where a common low-pass filter, such as a Gaussian filter, has a weight w (i, j, x, y) based on the distance from the center of the core (x, y) to each pixel (i, k). For a two-sided filter, the weight is determined based on two distances: the distance between the image space and the distance between the colors.

Simple C implementation below

void convolution(uchar4 *_in, uchar4 *_out, int width, int height, int ~ halfkernelsize, float id, float cd) { int kernelDim = 2*halfkernelsize+1; 
 for(int y=0; y float sumWeight = 0; unsigned int ctrIdx = y*width + x; float ctrPix[3]; ctrPix[0] = _in[ctrIdx].x; ctrPix[1] = _in[ctrIdx].y; ctrPix[2] = _in[ctrIdx].z; // neighborhood of current pixel int kernelStartX, kernelEndX, kernelStartY, kernelEndY; kernelStartX = x-halfkernelsize; kernelEndX = x+halfkernelsize; kernelStartY = y-halfkernelsize; kernelEndY = y+halfkernelsize; for(int j= kernelStartY; j<= kernelEndY; j++) { for(int i= kernelStartX; i<= kernelEndX; i++) { unsigned int idx = max(0, min(j, height-1))*width + max(0, min(i,width-1)); float curPix[3]; curPix[0] = _in[idx].x; curPix[1] = _in[idx].y; curPix[2] = _in[idx].z; float currWeight; // define bilateral filter kernel weights float imageDist = sqrt( (float)((ix)*(ix) + (jy)*(jy)) ); float colorDist = sqrt( (float)( (curPix[0] - ctrPix[0])*(curPix[0] - ctrPix[0]) + (curPix[1] - ctrPix[1])*(curPix[1] - ctrPix[1]) + (curPix[2] - ctrPix[2])*(curPix[2] - ctrPix[2]) ) ); currWeight = 1.0f/(exp((imageDist/id)*(imageDist/id)*0.5)*exp((colorDist/cd)*(colorDist/cd)*0.5)); sumWeight += currWeight; _sum[0] += currWeight*curPix[0]; _sum[1] += currWeight*curPix[1]; _sum[2] += currWeight*curPix[2]; } } _sum[0] /= sumWeight; _sum[1] /= sumWeight; _sum[2] /= sumWeight; _out[ctrIdx].x = (int)(floor(_sum[0])); _out[ctrIdx].y = (int)(floor(_sum[1])); _out[ctrIdx].z = (int)(floor(_sum[2])); _out[ctrIdx].w = _in[ctrIdx].w; } } 

}