1、OpenCV卷积函数filter2D():形式：CV_EXPORTS_W void filter2D( InputArray src,OutputArray dst,int ddepth, InputArray kernel,Point anchor=Point(-1,-1), double delta=0,int borderType=BORDER_DEFAULT );功能：利用内核实现对图像的卷积运算；参数：InputArray src: 输入图像；OutputArray dst: 输出图像，和输入图像具有相同的尺寸和通道数量；int ddepth: 目标图像深度；原图像和目标图像支持的图像深度如下：src.depth() = CV_8U, ddepth = -1/CV_16S/CV_32F/CV_64F src.depth() = CV_16U/CV_16S, ddepth = -1/CV_32F/CV_64F src.depth() = CV_32F, ddepth = -1/CV_32F/CV_64F src.depth() = CV_64F, ddepth = -1/CV_64F当ddepth输入值为-1时，目标图像和原图像深度保持一致；InputArray kernel: 卷积核,一个单通道浮点型矩阵。如果想在图像不同的通道使用不同的kernel，可以先使用split()函数将图像通道事先分开。Point anchor: 内核的基准点(anchor)，其默认值为(-1,-1)说明位于kernel的中心位置。基准点即kernel中与进行处理的像素点重合的点。double delta:在储存目标图像前可选的添加到像素的值，默认值为0int borderType:像素向外逼近的方法，默认值是BORDER_DEFAULT,即对全部边界进行计算。
2、filter2D()示例：#include <opencv2\opencv.hpp>using namespace std;using namespace cv;int main( int argc, char** argv ){ Mat raw_img = imread("raw.jpg",1); if (!raw_img.data) { return -1; } imshow("before",raw_img); Mat mask_img = (Mat_<char>(3,3) << 0,-1,0,-1,5,-1,0,-1,0); Mat dst_img(raw_img.rows,raw_img.cols,CV_32FC3); Point anchor = Point(-1,-1); filter2D(raw_img,dst_img,-1,mask_img,anchor,0,BORDER_DEFAULT); imshow("before",raw_img); imshow("after",dst_img); waitKey( 0 ); return 0;}
3、常见卷积模板：模板：矩阵方块，其数学含义是一种卷积运算；卷积运算：是加权求和的过程；所有乘积之和作为区域中心像素的新值；卷积示例：3*3的像素区域R与卷积核G的卷积运算(中心像素)R5=R1G1+R2G2+R3G3+R4G4+R5G5+R6G6+R7G7+R8G8+R9G9;常见模板如下：滤波是图像处理的基本操作，滤波去除图像中的噪声，提取感兴趣的特征，允许图像重采样。图像中的频域和空域：空间域指用图像的灰度值来描述一幅图像；而频域指用图像灰度值的变化来描述一幅图像。而低通滤波器和高通滤波器的概念就是在频域中产生的。低通滤波器：指去除图像中的高频成分，使边缘平滑；而高通滤波器：指去除图像中的低频成分，使边缘提取与增强；
4、卷积边界问题：当处理图像边界像素时，卷积核与图像使用区域不能匹配，卷积核的中心与边界像素点对应，卷积将出现问题。处理方法：A.忽略边界像素，即处理后的图像将丢掉这些像素；B.保留原边界像素，即copy边界像素到处理后的图像；
5、自定义卷积运算：#include <opencv2\opencv.hpp>using namespace std;using namespace cv;//kernel(-1,-2,1;4,-2,-1;4,-2,2);void Sharpen(Mat& myImage, Mat& Result){ int nChannels = myImage.channels(); Result.create(myImage.size(), myImage.type()); for (int j = 1; j < myImage.rows - 1; ++j) //hang { unsigned char *previous = myImageNaNr<unsigned char>(j-1);//上一行数据的指针 unsigned char *current = myImageNaNr<unsigned char>(j);//当前行数据的指针 unsigned char *next = myImageNaNr<unsigned char>(j+1);//下一行数据的指针 unsigned char *output = ResultNaNr<unsigned char>(j); //输出图像当前列数据的指针 for (int i = 1; i < myImage.cols - 1; ++i)//lie { output[i] += saturate_cast<uchar>( -1*previous[i-1]-2*previous[i]+1*previous[i+1] +4*current[i-1]-2*current[i]-1*current[i+1] +4*next[i-1]-2*next[i]+2*next[i+1]) ; } } // 将边界设为0 Result.row(0).setTo(Scalar(0)); Result.row(Result.rows - 1).setTo(Scalar(0)); Result.col(0).setTo(Scalar(0)); Result.col(Result.cols - 1).setTo(Scalar(0));}int main(){ Mat img = imread("raw.jpg",0); Mat src; Sharpen(img, src); imshow("img", img); imshow("src",src); waitKey(0); return 0;}
6、示例：#include <opencv2\opencv.hpp>using namespace std;using namespace cv;Mat Gaussian_kernal(int kernel_size,float sigma);Mat z_Sharpen(Mat raw_img,Mat kernel_img);int main( int argc, char** argv ){ Mat raw_img = imread("raw.jpg",0); imshow("before",raw_img); int kernel_size =3; float sigma = 1.4; Mat kernel_img = Gaussian_kernal(kernel_size,sigma); cout<<kernel_img<<endl; //float myArray[3][3] = //{ // -1, -2, 1, // 4, -2, -1, // 4, -2, 2 //}; //Mat kernel_img = Mat(3, 3, CV_32FC1, myArray);//创建矩阵 Mat dst_img = z_Sharpen(raw_img,kernel_img); imshow("after",dst_img); waitKey( 0 ); return 0;}Mat z_Sharpen(Mat raw_img,Mat kernel_img){ int m_width = raw_img.cols;//kuan int m_height = raw_img.rows;//gao int kernel_size = kernel_img.cols; int k_size = (kernel_size-1)/2; Mat dst_img(m_height,m_width,CV_8UC1); for (int j=k_size;j<m_width-k_size;j++)//lie { for (int i=k_size;i<m_height-k_size;i++)//hang { unsigned char *current = dst_imgNaNr<unsigned char>(i);//dang qian hang int m_J = -k_size;//lie for (int m=0;m<kernel_size;m++) { float *k_data = kernel_imgNaNr<float>(m); int m_I = -k_size;//hang for (int n=0;n<kernel_size;n++) { unsigned char *p_data = raw_imgNaNr<unsigned char>(i+m_I); current[j] += saturate_cast<unsigned char>(p_data[j+m_J]*k_data[n]); m_I++; } m_J++; } } } //将边界设为0 for (int i=0;i<k_size;i++) { dst_img.row(i).setTo(Scalar(0)); dst_img.col(i).setTo(Scalar(0)); } for(int i=m_height-k_size;i<m_height;i++) { dst_img.row(i).setTo(Scalar(0)); } for(int i=m_width-k_size;i<m_width;i++) { dst_img.col(i).setTo(Scalar(0)); } return dst_img;}Mat Gaussian_kernal(int kernel_size,float sigma){ const double PI = 3.14159265358979323846; int m = kernel_size / 2; Mat mask_temp(kernel_size,kernel_size,CV_32FC1); float s_2 = 2*sigma*sigma; for (int i=0;i<kernel_size;i++) { for (int j=0;j<kernel_size;j++) { int x=i-m; int y=j-m; mask_tempNaNr<float>(i)[j] = exp(-(x*x+y*y)/2*s_2) /(PI*s_2); } } return mask_temp;}
