图像/“最像像素”搜索优化?

假设我有一个图像A,比如说,512x512像素,和图像B,5x5或7x7像素。 两幅图像都是24位rgb,而B有1位alpha遮罩(因此每个像素要么是完全透明的,要么是完全实心的)。

我需要在图像A中找到一个与图像B最相似的像素,或者是 最像图像B。

相似度计算为“距离”,即非透明B的像素和A的像素之间的“距离”之和除以非透明B的像素数。以下是示例SDL代码以供解释:

struct Pixel{
    unsigned char b, g, r, a;
};

void fillPixel(int x, int y, SDL_Surface* dst, SDL_Surface* src, int dstMaskX, int dstMaskY){
    Pixel& dstPix = *((Pixel*)((char*)(dst->pixels) + sizeof(Pixel)*x + dst->pitch*y));

    int xMin = x + texWidth - searchWidth;
    int xMax = xMin + searchWidth*2;
    int yMin = y + texHeight - searchHeight;
    int yMax = yMin + searchHeight*2;


    int numFilled = 0;
    for (int curY = yMin; curY < yMax; curY++)
        for (int curX = xMin; curX < xMax; curX++){
            Pixel& cur = *((Pixel*)((char*)(dst->pixels) + sizeof(Pixel)*(curX & texMaskX) + dst->pitch*(curY & texMaskY)));
            if (cur.a != 0)
                numFilled++;
        }

    if (numFilled == 0){
        int srcX = rand() % src->w;
        int srcY = rand() % src->h;
        dstPix = *((Pixel*)((char*)(src->pixels) + sizeof(Pixel)*srcX + src->pitch*srcY));
        dstPix.a = 0xFF;
        return;
    }

    int storedSrcX = rand() % src->w;
    int storedSrcY = rand() % src->h;
    float lastDifference = 3.40282347e+37F;

    //unsigned char mask = 

    for (int srcY = searchHeight; srcY < (src->h - searchHeight); srcY++)
        for (int srcX = searchWidth; srcX < (src->w - searchWidth); srcX++){
            float curDifference = 0;
            int numPixels = 0;
            for (int tmpY = -searchHeight; tmpY < searchHeight; tmpY++)
                for(int tmpX = -searchWidth; tmpX < searchWidth; tmpX++){
                    Pixel& tmpSrc = *((Pixel*)((char*)(src->pixels) + sizeof(Pixel)*(srcX+tmpX) + src->pitch*(srcY+tmpY)));
                    Pixel& tmpDst = *((Pixel*)((char*)(dst->pixels) + sizeof(Pixel)*((x + dst->w + tmpX) & dstMaskX) + dst->pitch*((y + dst->h + tmpY) & dstMaskY)));
                    if (tmpDst.a){
                        numPixels++;
                        int dr = tmpSrc.r - tmpDst.r;
                        int dg = tmpSrc.g - tmpDst.g;
                        int db = tmpSrc.g - tmpDst.g;
                        curDifference += dr*dr + dg*dg + db*db;
                    }
                }
            if (numPixels)
                curDifference /= (float)numPixels;
            if (curDifference < lastDifference){
                lastDifference = curDifference;
                storedSrcX = srcX;
                storedSrcY = srcY;
            }
        }

    dstPix = *((Pixel*)((char*)(src->pixels) + sizeof(Pixel)*storedSrcX + src->pitch*storedSrcY));
    dstPix.a = 0xFF;
}

最简单的方法是暴力搜索(在示例例程中使用)。但它是缓慢的-即使使用GPU加速和双核cpu不会使它更快。看来我不能使用修改过的二进制搜索,因为B的掩码。那么,怎样才能更快地找到所需的像素呢?

其他信息:

1. 它允许使用2个核心,GPU加速,CUDA和1.5..2 GB的RAM来完成任务。
2. 我宁愿避免一些冗长的预处理阶段,这将需要30分钟才能完成。

思想?