标签：
图像平滑是图像预处理的基本操作,本文首先用不同的方法对一张图片做预处理比较它们效果的不同,然后针对中值滤波,实现了一种快速实现.(其实是copy的opencv实现,呵呵).因为opencv的实现方法感觉太好了,今天就特别写下来.既有备忘的作用,同时如果谁看到这篇文章,也可以借鉴下opencv的实现.
| void CDialogTest2013Dlg::OnBnClickedMfcbutton1() { char file[10] = "e:\\1.jpg";
IplImage *src = cvLoadImage(file, 0); CvSize picSize; picSize.width = src->width; picSize.height = src->height; IplImage *dst_midian= cvCreateImage(picSize, 8, 1); IplImage *dst_blur = cvCreateImage(picSize, 8, 1); IplImage *dst_gaussian = cvCreateImage(picSize, 8, 1); IplImage *dst_bilateral = cvCreateImage(picSize, 8, 1); cvSmooth(src, dst_midian, CV_MEDIAN, 3, 3); cvSmooth(src, dst_blur, CV_BLUR, 3, 3); cvSmooth(src, dst_gaussian, CV_GAUSSIAN, 3, 3); cvSmooth(src, dst_bilateral, CV_BILATERAL, 3, 3);
} |
均值滤波原理:把一个点周围点求均值,这个均值给这个点.
中值滤波原理:把一个点周围的点排序,把中位数给这个点.
双边滤波原理:这个一句话说不清楚,可以看这个博文 http://blog.csdn.net/bugrunner/article/details/7170471
高斯滤波原理:均值滤波对一个点周围点求和是每个点的权重都是1,高斯滤波的权重不是1了,是一个类似高斯分布的权重分布,距离目标点
近的地方权重高点,距离目标点远的地方,权重低些.
中值滤波的opencv实现
| template<class Op, class VecOp> static void medianBlur_SortNet( const Mat& _src, Mat& _dst, int m ) { typedef typename Op::value_type T; typedef typename Op::arg_type WT; typedef typename VecOp::arg_type VT; const T* src = (const T*)_src.data; T* dst = (T*)_dst.data; int sstep = (int)(_src.step/sizeof(T)); int dstep = (int)(_dst.step/sizeof(T)); Size size = _dst.size(); int i, j, k, cn = _src.channels(); Op op; VecOp vop; volatile bool useSIMD = checkHardwareSupport(CV_CPU_SSE2); if( m == 3 ) { if( size.width == 1 || size.height == 1 ) { int len = size.width + size.height - 1; int sdelta = size.height == 1 ? cn : sstep; int sdelta0 = size.height == 1 ? 0 : sstep - cn; int ddelta = size.height == 1 ? cn : dstep; for( i = 0; i < len; i++, src += sdelta0, dst += ddelta ) for( j = 0; j < cn; j++, src++ ) { WT p0 = src[i > 0 ? -sdelta : 0]; WT p1 = src[0]; WT p2 = src[i < len - 1 ? sdelta : 0]; op(p0, p1); op(p1, p2); op(p0, p1); dst[j] = (T)p1; } return; } size.width *= cn; for( i = 0; i < size.height; i++, dst += dstep ) { const T* row0 = src + std::max(i - 1, 0)*sstep; const T* row1 = src + i*sstep; const T* row2 = src + std::min(i + 1, size.height-1)*sstep; int limit = useSIMD ? cn : size.width; for(j = 0;; ) { for( ; j < limit; j++ ) { int j0 = j >= cn ? j - cn : j; int j2 = j < size.width - cn ? j + cn : j; WT p0 = row0[j0], p1 = row0[j], p2 = row0[j2]; WT p3 = row1[j0], p4 = row1[j], p5 = row1[j2]; WT p6 = row2[j0], p7 = row2[j], p8 = row2[j2]; op(p1, p2); op(p4, p5); op(p7, p8); op(p0, p1); //op(p1, p2) 交换p1和p2两个数 op(p3, p4); op(p6, p7); op(p1, p2); op(p4, p5); op(p7, p8); op(p0, p3); op(p5, p8); op(p4, p7); op(p3, p6); op(p1, p4); op(p2, p5); op(p4, p7); op(p4, p2); op(p6, p4); op(p4, p2); dst[j] = (T)p4; } if( limit == size.width ) break; for( ; j <= size.width - VecOp::SIZE - cn; j += VecOp::SIZE ) { VT p0 = vop.load(row0+j-cn), p1 = vop.load(row0+j), p2 = vop.load(row0+j+cn); VT p3 = vop.load(row1+j-cn), p4 = vop.load(row1+j), p5 = vop.load(row1+j+cn); VT p6 = vop.load(row2+j-cn), p7 = vop.load(row2+j), p8 = vop.load(row2+j+cn); vop(p1, p2); vop(p4, p5); vop(p7, p8); vop(p0, p1); vop(p3, p4); vop(p6, p7); vop(p1, p2); vop(p4, p5); vop(p7, p8); vop(p0, p3); vop(p5, p8); vop(p4, p7); vop(p3, p6); vop(p1, p4); vop(p2, p5); vop(p4, p7); vop(p4, p2); vop(p6, p4); vop(p4, p2); vop.store(dst+j, p4); } limit = size.width; } } } } }
|
标签：
原文地址：http://www.cnblogs.com/guopengfei/p/4745676.html
