自动聚焦
========================================

实验介绍
========================================

相机的自动对焦要求相机根据拍摄环境和场景的变化,通过相机内部的微型驱动马达,自动调节相机镜头和CMOS之间的距离,保证像平面正好投影到CMOS的成像表面上。这时候物体的成像比较清晰,图像细节信息丰富。
图像清晰度评价算法有很多种,在空域中,主要思路是考察图像的领域对比度,即相邻像素间的灰度特征的梯度差；在频域中,主要思路是考察图像的频率分量,对焦清晰的图像高频分量较多,对焦模糊的图像低频分量较多。
本实验选用Tenengrad梯度方法,利用Sobel算子分别计算水平和垂直方向的梯度,同一场景下梯度值越高,图像越清晰。
我们将焦点由远及近调整,并比较各个焦点位置的Sobel结果,值越大则表示该点离最佳聚焦点越近。
因为有的镜头不支持自动调整焦点,对于不支持调焦的镜头,需要手动调整焦点,通过评估值的显示,来确定最佳焦点。

代码
========================================

.. code:: c

  #include "sobel_focus.h"

  void ExtractPixel(XF_TNAME(XF_8UC3,XF_NPPC1)&src,ap_uint<8>src_value[3])
  {
  #pragma HLS INLINE off
  	unsigned int i,j=0;
  	for(i=0;i<24;i+=8)
  	{
  #pragma HLS UNROLL
  		src_value[j]=src.range(i+7,i);
  		j++;
  	}
  }
  
  template<int ROWS, int COLS>
  void sum_of_stream(xf::cv::Mat<XF_8UC1,ROWS,COLS,XF_NPPC1>	&_src, unsigned int &average)
  {
  	XF_TNAME(XF_8UC1,XF_NPPC1)pix;
  	unsigned int sum=0;
  
  	for(int v=0; v<ROWS; v++)
  	{
  		for(int h=0; h<COLS; h++)
  		{
  #pragma HLS pipeline
  			pix = _src.read(v*COLS+h);
  			sum += pix.range(7,0);
  		}
  	}
  	average = sum;
  }
  
  template<int ROWS,int COLS>
  void xfrgb2gray(xf::cv::Mat<XF_8UC3,ROWS,COLS,XF_NPPC1> &src,xf::cv::Mat<XF_8UC1,ROWS,COLS,XF_NPPC1> &dst)
  {
  	XF_TNAME(XF_8UC3,XF_NPPC1)rgb_packed;
  	XF_TNAME(XF_8UC1,XF_NPPC1)gray_packed;
  	ap_uint<8>rgb[3];
  	ap_uint<8>gray;
  	unsigned int i,j=0;
  	for(i=0;i<ROWS;i++)
  	{
  		for(j=0;j<COLS;j++)
  		{
  #pragma HLS PIPELINE
  			rgb_packed=src.read(i*COLS+j);
  			ExtractPixel(rgb_packed,rgb);
  			gray=CalculateGRAY(rgb[0],rgb[1],rgb[2]);
  			gray_packed.range(7,0)=gray;
  			dst.write(i*COLS+j,gray_packed);
  		}
  	}
  }
  
  template<int ROWS,int COLS>
  void AddWeightedKernel(xf::cv::Mat<XF_8UC1,ROWS,COLS,XF_NPPC1>&src1,
  					   float alpha,
  					   xf::cv::Mat<XF_8UC1,ROWS,COLS,XF_NPPC1>&src2,
  					   float beta,
  					   float gamma,
  					   xf::cv::Mat<XF_8UC1,ROWS,COLS,XF_NPPC1>&dst
  					)
  {
  	ap_fixed<16,8,AP_RND>value_src1=alpha;
  	ap_fixed<16,8,AP_RND>value_src2=beta;
  	ap_fixed<16,8,AP_RND>value_src3=gamma;
  	XF_TNAME(XF_8UC1,XF_NPPC1)pixel1;
  	XF_TNAME(XF_8UC1,XF_NPPC1)pixel2;
  	XF_TNAME(XF_8UC1,XF_NPPC1)pixel3;
  	ap_int<24>firstcmp;
  	ap_int<24>secondcmp;
  	ap_int<16>thirdcmp;
  	ap_uint<8>value;
  	ap_uint<8>value_cmp1;
  	ap_uint<8>value_cmp2;
  	unsigned int i,j=0;
  	for(i=0;i<ROWS;i++)
  	{
  		for(j=0;j<COLS;j++)
  		{
  #pragma HLS pipeline
  			pixel1=src1.read(i*COLS+j);
  			pixel2=src2.read(i*COLS+j);
  			value_cmp1=pixel1.range(7,0);
  			value_cmp2=pixel2.range(7,0);
  			firstcmp=(ap_int<24>)value_cmp1*value_src1;
  			secondcmp=(ap_int<24>)value_cmp2*value_src2;
  			thirdcmp=(ap_int<16>)firstcmp+secondcmp+value_src3;
  			if(thirdcmp>255)
  			{
  				thirdcmp=255;
  			}
  			else if(thirdcmp<0)
  			{
  				thirdcmp=0;
  			}
  			value=thirdcmp;
  			pixel3.range(7,0)=value;
  			dst.write(i*COLS+j,pixel3);
  		}
  	}
  }
  
  template<int ROWS,int COLS>
  void duplicate(xf::cv::Mat<XF_8UC3,ROWS,COLS,XF_NPPC1>&src,xf::cv::Mat<XF_8UC3,ROWS,COLS,XF_NPPC1>&dst1,xf::cv::Mat<XF_8UC3,ROWS,COLS,XF_NPPC1>&dst2)
  {
  	unsigned int i,j=0;
  	XF_TNAME(XF_8UC3,XF_NPPC1)pixel_src;
  	for(i=0;i<ROWS;i++)
  	{
  		for(j=0;j<COLS;j++)
  		{
  #pragma HLS PIPELINE
  			pixel_src=src.read(i*COLS+j);
  			dst1.write(i*COLS+j,pixel_src);
  			dst2.write(i*COLS+j,pixel_src);
  		}
  	}
  }
  
  
  void sobel_focus(ustream_t &src, ustream_t &dst, unsigned int &average)
  {
  #pragma HLS INTERFACE axis port=src
  #pragma HLS INTERFACE axis port=dst
  #pragma HLS INTERFACE s_axilite port=average
  #pragma HLS INTERFACE ap_ctrl_none port=return
  
  	xf::cv::Mat<XF_8UC3,IMG_MAX_ROWS,IMG_MAX_COLS,XF_NPPC1> srcImg, split0, split1;
  #pragma HLS STREAM depth=1920 type=fifo variable=split1
  #pragma HLS STREAM depth=1920 type=fifo variable=split0
  #pragma HLS STREAM depth=1920 type=fifo variable=srcImg
  	xf::cv::Mat<XF_8UC1,IMG_MAX_ROWS,IMG_MAX_COLS,XF_NPPC1>grayImg,sobelImg_x,sobelImg_y,sobelImg;
  #pragma HLS STREAM depth=1920 type=fifo variable=sobelImg
  #pragma HLS STREAM depth=1920 type=fifo variable=sobelImg_y
  #pragma HLS STREAM depth=1920 type=fifo variable=sobelImg_x
  #pragma HLS STREAM depth=1920 type=fifo variable=grayImg
  #pragma HLS DATAFLOW
  	xf::cv::AXIvideo2xfMat(src, srcImg);
  	duplicate<IMG_MAX_ROWS,IMG_MAX_COLS>(srcImg, split0, split1);
  	xfrgb2gray<IMG_MAX_ROWS,IMG_MAX_COLS>(split0, grayImg);
      xf::cv::xFSobelFilter3x3<XF_8UC1, XF_8UC1,IMG_MAX_ROWS, IMG_MAX_COLS, XF_CHANNELS(XF_8UC1,XF_NPPC1), XF_DEPTH(XF_8UC1,XF_NPPC1), XF_DEPTH(XF_8UC1,XF_NPPC1),
                      XF_NPPC1, _XFCVDEPTH_DEFAULT,_XFCVDEPTH_DEFAULT,_XFCVDEPTH_DEFAULT,XF_WORDWIDTH(XF_8UC1,XF_NPPC1), XF_WORDWIDTH(XF_8UC1,XF_NPPC1), (IMG_MAX_COLS >> XF_BITSHIFT(XF_NPPC1)),false>(
          grayImg,sobelImg_x,sobelImg_y,grayImg.rows,grayImg.cols>>XF_BITSHIFT(XF_NPPC1));
  	AddWeightedKernel<IMG_MAX_ROWS,IMG_MAX_COLS>(sobelImg_x,0.5f,sobelImg_y,0.5f,0.0f,sobelImg);
  	sum_of_stream(sobelImg, average);
  	xf::cv::xfMat2AXIvideo(split1, dst);
  }

工程路径
==========================================

.. csv-table:: 
  :header: "名称", "路径"
  :widths: 20, 20

  "vivado 工程","vivado/auto_focus"
  "HLS工程","vivado/sobel_focus"
  "HLS工程","hls/mem2stream"
  "HLS工程","hls/stream2mem"
  "BOOT.bin文件","bootimage"

实验结果
===========================================
默认焦点在最远处

.. image:: images/images6/image60.png
      
焦点拉近至最佳点

.. image:: images/images6/image61.png