我的SSE图像优化算法2.图像二值化

自己写的C++源码性能远不如SSE代码，每次实现也很麻烦，所以直接将SSE与opencv的库进行对比，opencv虽然也远不如商用视觉算法库，但也算是优化的很好了，下面看一下他和SSE的速度差别吧！
以下分别是单阈值和双阈值的SSE代码：

void ThresholdBinarize(cv::Mat InImg, cv::Mat& OutImg, unsigned char nThreshold)
{
	if (InImg.type() != CV_8UC1)
	{
		return ;
	}
	OutImg = Mat(InImg.rows, InImg.cols, CV_8UC1);
	unsigned char *Src = InImg.data;
	unsigned char *Dst = OutImg.data;
	int Width = InImg.cols;
	int Height = InImg.rows;
	const int BlockSize = 16;
	int Block = Width / BlockSize;
	for (int Y = 0; Y < Height; Y++)
	{
		unsigned char *LinePS = Src + Y * Width;
		unsigned char *LinePD = Dst + Y * Width;
		for (int X = 0; X < Block * BlockSize; X += BlockSize)
		{
			__m128i Src1, Result;
			Src1 = _mm_loadu_si128((__m128i *)(LinePS + X));
			_mm_set1_epi8(nThreshold));
			Result = _mm_cmpeq_epi8(_mm_max_epu8(Src1, _mm_set1_epi8(nThreshold)), Src1);
			_mm_storeu_si128((__m128i*)(LinePD + X), Result);
		}
		for (int X = Block * BlockSize; X < Width; X++)
		{
			LinePD[X] = LinePS[X] >= nThreshold ? 255 : 0;
		}
	}
}

void ThresholdBinarize(cv::Mat InImg, cv::Mat& OutImg, unsigned char nMinThreshold, unsigned char nMaxThreshold) {
	if (InImg.type() != CV_8UC1)
	{
		return;
	}
	OutImg = Mat(InImg.rows, InImg.cols, CV_8UC1);
	unsigned char *Src = InImg.data;
	unsigned char *Dst = OutImg.data;
	int Width = InImg.cols;
	int Height = InImg.rows;
	const int BlockSize = 16;
	int Block = Width / BlockSize;
	for (int Y = 0; Y < Height; Y++)
	{
		unsigned char *LinePS = Src + Y * Width;
		unsigned char *LinePD = Dst + Y * Width;
		for (int X = 0; X < Block * BlockSize; X += BlockSize)
		{
			__m128i Src1, Result;
			Src1 = _mm_loadu_si128((__m128i *)(LinePS + X));
			Result = _mm_cmpeq_epi8(_mm_max_epu8(Src1, _mm_set1_epi8(nMinThreshold)), Src1);
			Result = _mm_and_si128(Result, _mm_cmpeq_epi8(_mm_min_epu8(Src1, _mm_set1_epi8(nMaxThreshold)), Src1));
			_mm_storeu_si128((__m128i*)(LinePD + X), Result);
		}
		for (int X = Block * BlockSize; X < Width; X++)
		{
			LinePD[X] = LinePS[X] >= nMinThreshold && LinePS[X] <= nMaxThreshold ? 255 : 0;
		}
	}
}

二值化图像有一个判断的操作：

if (LinePD[S]<=nThreshold)
	LinePD[X] = 255; 

判断指令在SSE是很难实现的，我这里使用了

_mm_cmpeq_epi8(_mm_max_epu8(Src1, _mm_set1_epi8(nThreshold)), Src1);

_mm_max_epu8(Src1, _mm_set1_epi8(nThreshold))将Src1与nThreshold比较并取较大的值，Src1中超过nThreshold的值便能不变，而小于nThreshold的值置为nThreshold。
再通过_mm_cmpeq_epi8将该值与Src1做比较,如果相等，则令结果等于255，如果不相等，则令结果等于0，这样便能使原图中超过nThreshold的值置为255，小于nThreshold的值置为0。

opencv直接调用threshold(srcImg, dst, 50, 255, CV_THRESH_BINARY);
结果如下：

需要注意的是双阈值的SSE实现功能与threshold(srcImg, dst, 50, 255, CV_THRESH_BINARY)并不相同。SSE的功能是将灰度大于小阈值并小于大阈值的像素置为255，其他像素置为0。而opencv的threshold是将灰度大于第一个参数的区域的像素置为第二个参数。

虽然达不到理论上的速度，但SSE还是明显快于opencv，毕竟opencv也是优化的很好了，而我的SSE代码没有进行过多的优化。那咋办嘛，凑合着过呗。