Warning: file_get_contents(/data/phpspider/zhask/data//catemap/0/drupal/3.json): failed to open stream: No such file or directory in /data/phpspider/zhask/libs/function.php on line 167

Warning: Invalid argument supplied for foreach() in /data/phpspider/zhask/libs/tag.function.php on line 1116

Notice: Undefined index: in /data/phpspider/zhask/libs/function.php on line 180

Warning: array_chunk() expects parameter 1 to be array, null given in /data/phpspider/zhask/libs/function.php on line 181
Image 基于CUDA和OpenCV的Sobel边缘检测在灰度jpg图像上的应用_Image_Opencv_Cuda - Fatal编程技术网

Image 基于CUDA和OpenCV的Sobel边缘检测在灰度jpg图像上的应用

Image 基于CUDA和OpenCV的Sobel边缘检测在灰度jpg图像上的应用,image,opencv,cuda,Image,Opencv,Cuda,这个问题以前已经被问过了,但是提问者没有提供足够的信息,没有回答,我对这个项目很好奇 我正在尝试使用opencv和cuda库进行sobel边缘检测, X方向的sobel核是 -1 0 1 -2 0 2 -1 0 1 我的项目中有3个文件 main.cpp CudaKernel.cu CudaKernel.h main.cpp #include <stdlib.h> #include <iostream> #include <string.h&g

这个问题以前已经被问过了,但是提问者没有提供足够的信息,没有回答,我对这个项目很好奇

我正在尝试使用opencv和cuda库进行sobel边缘检测, X方向的sobel核是

-1 0 1   
-2 0 2  
-1 0 1   
我的项目中有3个文件

main.cpp
CudaKernel.cu
CudaKernel.h
main.cpp

#include <stdlib.h>
#include <iostream>
#include <string.h>
#include <Windows.h>
#include <opencv2\core\core.hpp>
#include <opencv2\highgui\highgui.hpp>
#include <opencv2\gpu\gpu.hpp>
#include <cuda_runtime.h>
#include <cuda_gl_interop.h>
#include "CudaKernel.h"

using namespace cv;
using namespace std;


int main(int argc, char** argv) 
{
    IplImage* image;

    try
    {
        image = cvLoadImage("4555472_460s.jpg", CV_LOAD_IMAGE_GRAYSCALE);
        gpu::DeviceInfo info = gpu::getDevice();
        cout << info.name() << endl;
        cout << "Stream Processor : "<< info.multiProcessorCount() << endl;
        cout << "Total Graphic Memory :" << info.totalMemory()/1048576 << " MB" << endl; 
    }
    catch (const cv::Exception* ex)
    {
        cout << "Error: " << ex->what() << endl;
    }
    if(!image )
        {
             cout << "Could not open or find the image" << std::endl ;
             return -1;
        }


    IplImage* image2=cvCreateImage(cvGetSize(image),IPL_DEPTH_32F,image->nChannels);
    IplImage* image3=cvCreateImage(cvGetSize(image),IPL_DEPTH_32F,image->nChannels);

    unsigned char * pseudo_input=(unsigned char *)image->imageData;
    float *output=(float*)image2->imageData;
    float *input=(float*)image3->imageData;
    int s=image->widthStep/sizeof(float);
        for(int w=0;w<=(image->height);w++)
            for(int h=0;h<(image->width*image->nChannels);h++)
            {
                input[w*s+h]= pseudo_input[w*s+h];
            }


    Pixel *fagget  = (unsigned char*) image->imageData;
    kernelcall(input, output, image->width,image->height, image->widthStep);

//  cv::namedWindow( "Display window", CV_WINDOW_AUTOSIZE );// Create a window for display.
    cvShowImage( "Original Image", image ); // Show our image inside it.
    cvShowImage("Sobeled Image", image2);
    waitKey(0); // Wait for a keystroke in the window
    return 0;

}
#include <iostream>
#include <opencv2/opencv.hpp>
#include "CudaKernel.h"

using namespace cv;
using namespace std;

int main(int argc, char** argv) 
{
    IplImage* image;

    image = cvLoadImage("4555472_460s.jpg", CV_LOAD_IMAGE_GRAYSCALE);

    if(!image )
    {
        cout << "Could not open or find the image" << std::endl;
        return -1;
    }


    IplImage* image2 = cvCreateImage(cvGetSize(image),IPL_DEPTH_32F,image->nChannels);
    IplImage* image3 = cvCreateImage(cvGetSize(image),IPL_DEPTH_32F,image->nChannels);

    //Convert the input image to float
    cvConvert(image,image3);

    float *output = (float*)image2->imageData;
    float *input =  (float*)image3->imageData;

    kernelcall(input, output, image->width,image->height, image3->widthStep);

    //Normalize the output values from 0.0 to 1.0
    cvScale(image2,image2,1.0/255.0);

    cvShowImage("Original Image", image );
    cvShowImage("Sobeled Image", image2);
    cvWaitKey(0);
    return 0;
}
#包括
#包括
#包括
#包括
#包括
#包括
#包括
#包括
#包括
#包括“CudaKernel.h”
使用名称空间cv;
使用名称空间std;
int main(int argc,字符**argv)
{
IplImage*图像;
尝试
{
image=cvLoadImage(“4555472_460s.jpg”,CV_LOAD_image_灰度);
gpu::DeviceInfo info=gpu::getDevice();
cout图像数据;
浮点*输入=(浮点*)图像3->图像数据;
float*inputnormalized=(float*)image4->imageData;
int s=图像->宽度步长/sizeof(浮点);
for(int w=0;wheight);w++)
对于(int h=0;hwidth*image->n通道);h++)
{
输入[w*s+h]=伪_输入[w*s+h];
}
内核调用(输入、输出、图像->宽度、图像->高度、图像->宽度步长);
cvNormalize(输入、输入规范化、0255、标准最小值、CV 8UC1);
cvShowImage(“原始图像”,Image);//在其中显示我们的图像。
cvShowImage(“Sobeled图像”,图像2);
但是现在我得到一个未处理的异常错误。

这里:-
Here:-

unsigned char * pseudo_input=(unsigned char *)image->imageData;
float *output=(float*)image2->imageData;
float *input=(float*)image3->imageData;
int s=image->widthStep/sizeof(float);
    for(int w=0;w<=(image->height);w++)
        for(int h=0;h<(image->width*image->nChannels);h++)
        {
            input[w*s+h]= pseudo_input[w*s+h];
        }
无符号字符*伪_输入=(无符号字符*)图像->图像数据; 浮点*输出=(浮点*)图像2->图像数据; 浮点*输入=(浮点*)图像3->图像数据; int s=图像->宽度步长/sizeof(浮点); for(int w=0;wheight);w++) 对于(int h=0;hwidth*image->n通道);h++) { 输入[w*s+h]=伪_输入[w*s+h]; }
输入为float*,伪输入为uchar*。将所有内容转换为float,然后进行处理。最后使用cvNormalize with NORM_MINMAX在0和255之间进行规格化,以获得正确的结果。

OpenCV规则编号1:

切勿通过基础数据指针直接访问图像数据,除非 绝对必要,例如将数据复制到GPU。参考(Me:p)

错误/建议:

  • 而不是通过循环图像数据来转换图像 指针,使用
    cvConvert
    更改图像数据类型。循环非常有用 很容易出错

  • 调用名为
    kernelcall
    的函数时,传递的是
    float
    图像的数据指针,但传递 原始8位图像。这是导致以下错误结果的主要原因: 这将导致内核内的索引错误

  • 在两个倾斜指针之间执行内存复制时 不同的步骤,始终使用可用的2D内存复制功能 在CUDA运行时,例如
    cudaMemcpy2D
    cudaMemcpy2DToArray
    等。在您的情况下,
    cuArray
    内部具有未知的宽度步长,并且输入的
    IplImage
    cuArray
    具有不同的宽度步长

  • 避免不必要的标题、赋值和标识符声明

  • 在CUDA内核中添加绑定检查,以便只有那些线程执行映像中的内存读/写。这可能会导致一些差异,但比无效内存读/写要好

  • 修订代码(已测试): Main.cpp

    #include <stdlib.h>
    #include <iostream>
    #include <string.h>
    #include <Windows.h>
    #include <opencv2\core\core.hpp>
    #include <opencv2\highgui\highgui.hpp>
    #include <opencv2\gpu\gpu.hpp>
    #include <cuda_runtime.h>
    #include <cuda_gl_interop.h>
    #include "CudaKernel.h"
    
    using namespace cv;
    using namespace std;
    
    
    int main(int argc, char** argv) 
    {
        IplImage* image;
    
        try
        {
            image = cvLoadImage("4555472_460s.jpg", CV_LOAD_IMAGE_GRAYSCALE);
            gpu::DeviceInfo info = gpu::getDevice();
            cout << info.name() << endl;
            cout << "Stream Processor : "<< info.multiProcessorCount() << endl;
            cout << "Total Graphic Memory :" << info.totalMemory()/1048576 << " MB" << endl; 
        }
        catch (const cv::Exception* ex)
        {
            cout << "Error: " << ex->what() << endl;
        }
        if(!image )
            {
                 cout << "Could not open or find the image" << std::endl ;
                 return -1;
            }
    
    
        IplImage* image2=cvCreateImage(cvGetSize(image),IPL_DEPTH_32F,image->nChannels);
        IplImage* image3=cvCreateImage(cvGetSize(image),IPL_DEPTH_32F,image->nChannels);
    
        unsigned char * pseudo_input=(unsigned char *)image->imageData;
        float *output=(float*)image2->imageData;
        float *input=(float*)image3->imageData;
        int s=image->widthStep/sizeof(float);
            for(int w=0;w<=(image->height);w++)
                for(int h=0;h<(image->width*image->nChannels);h++)
                {
                    input[w*s+h]= pseudo_input[w*s+h];
                }
    
    
        Pixel *fagget  = (unsigned char*) image->imageData;
        kernelcall(input, output, image->width,image->height, image->widthStep);
    
    //  cv::namedWindow( "Display window", CV_WINDOW_AUTOSIZE );// Create a window for display.
        cvShowImage( "Original Image", image ); // Show our image inside it.
        cvShowImage("Sobeled Image", image2);
        waitKey(0); // Wait for a keystroke in the window
        return 0;
    
    }
    
    #include <iostream>
    #include <opencv2/opencv.hpp>
    #include "CudaKernel.h"
    
    using namespace cv;
    using namespace std;
    
    int main(int argc, char** argv) 
    {
        IplImage* image;
    
        image = cvLoadImage("4555472_460s.jpg", CV_LOAD_IMAGE_GRAYSCALE);
    
        if(!image )
        {
            cout << "Could not open or find the image" << std::endl;
            return -1;
        }
    
    
        IplImage* image2 = cvCreateImage(cvGetSize(image),IPL_DEPTH_32F,image->nChannels);
        IplImage* image3 = cvCreateImage(cvGetSize(image),IPL_DEPTH_32F,image->nChannels);
    
        //Convert the input image to float
        cvConvert(image,image3);
    
        float *output = (float*)image2->imageData;
        float *input =  (float*)image3->imageData;
    
        kernelcall(input, output, image->width,image->height, image3->widthStep);
    
        //Normalize the output values from 0.0 to 1.0
        cvScale(image2,image2,1.0/255.0);
    
        cvShowImage("Original Image", image );
        cvShowImage("Sobeled Image", image2);
        cvWaitKey(0);
        return 0;
    }
    
    #包括
    #包括
    #包括“CudaKernel.h”
    使用名称空间cv;
    使用名称空间std;
    int main(int argc,字符**argv)
    {
    IplImage*图像;
    image=cvLoadImage(“4555472_460s.jpg”,CV_LOAD_image_灰度);
    如果(!图像)
    {
    壁炉);
    //将输入图像转换为浮点
    cvConvert(图像,图像3);
    浮点*输出=(浮点*)图像2->图像数据;
    浮点*输入=(浮点*)图像3->图像数据;
    内核调用(输入,输出,图像->宽度,图像->高度,图像3->宽度步长);
    //将输出值从0.0规范化为1.0
    cvScale(图像2,图像2,1.0/255.0);
    cvShowImage(“原始图像”,图像);
    cvShowImage(“Sobeled图像”,图像2);
    cvWaitKey(0);
    返回0;
    }
    
    CudaKernel.cu

    #include<cuda.h>
    #include<iostream>
    #include "CudaKernel.h"
    using namespace std;
    #define CudaSafeCall( err ) __cudaSafeCall( err, __FILE__, __LINE__ )
    #define CudaCheckError()    __cudaCheckError( __FILE__, __LINE__ )
    #define checkCudaErrors(err) __checkCudaErrors (err, __FILE__, __LINE__)
    
    
    texture <float,2,cudaReadModeElementType> tex1;
    texture<unsigned char, 2> tex;
    static cudaArray *array = NULL;
    static cudaArray *cuArray = NULL;
    
    
    //Kernel for x direction sobel
    __global__ void implement_x_sobel(float* garbage,float* output,int width,int height,int widthStep)
    {
        int x=blockIdx.x*blockDim.x+threadIdx.x;
        int y=blockIdx.y*blockDim.y+threadIdx.y;
    
        float output_value=((0*tex2D(tex1,x,y))+(2*tex2D(tex1,x+1,y))+(-2*tex2D(tex1,x-  1,y))+(0*tex2D(tex1,x,y+1))+(1*tex2D(tex1,x+1,y+1))+(-1*tex2D(tex1,x-1,y+1))+  (1*tex2D(tex1,x+1,y-1))+(0*tex2D(tex1,x,y-1))+(-1*tex2D(tex1,x-1,y-1)));
        output[y*widthStep+x]=output_value;
    }
    
    
    inline void __checkCudaErrors( cudaError err, const char *file, const int line )
    {
        if( cudaSuccess != err) {
            fprintf(stderr, "%s(%i) : CUDA Runtime API error %d: %s.\n",
                    file, line, (int)err, cudaGetErrorString( err ) );
            exit(-1);
        }
    }   
    
    //Host Code
     inline void __cudaSafeCall( cudaError err, const char *file, const int line )
    {
    #ifdef CUDA_ERROR_CHECK
    if ( cudaSuccess != err )
    {
        printf("cudaSafeCall() failed at %s:%i : %s\n",
                 file, line, cudaGetErrorString( err ) );
        exit( -1 );
    }    
    #endif
    
    return;
    }
    inline void __cudaCheckError( const char *file, const int line )
    {
    #ifdef CUDA_ERROR_CHECK
    cudaError err = cudaGetLastError();
    if ( cudaSuccess != err )
    {
        printf("cudaCheckError() failed at %s:%i : %s\n",
                 file, line, cudaGetErrorString( err ) );
       exit( -1 );
    }
    #endif
    
    return;
    }
    
    void kernelcall(float* input,float* output,int width,int height,int widthStep){
        //cudaChannelFormatDesc channelDesc=cudaCreateChannelDesc(32,32,0,0,cudaChannelFormatKindFloat);
        cudaChannelFormatDesc channelDesc = cudaCreateChannelDesc<float>();
        //cudaArray *cuArray;
        CudaSafeCall(cudaMallocArray(&cuArray,&channelDesc,width,height));
        cudaMemcpyToArray(cuArray,0,0,input,widthStep*height,cudaMemcpyHostToDevice);
    
        tex1.addressMode[0]=cudaAddressModeClamp;
        tex1.addressMode[1]=cudaAddressModeClamp;
        tex1.filterMode=cudaFilterModeLinear;
        cudaBindTextureToArray(tex1,cuArray,channelDesc);
        tex1.normalized=false;
        float * D_output_x;
        float * garbage=NULL;
        CudaSafeCall(cudaMalloc(&D_output_x,widthStep*height)); 
        dim3 blocksize(16,16);
        dim3 gridsize;
        gridsize.x=(width+blocksize.x-1)/blocksize.x;
        gridsize.y=(height+blocksize.y-1)/blocksize.y;
    
        implement_x_sobel<<<gridsize,blocksize>>>(garbage,D_output_x,width,height,widthStep/sizeof(float));
        cudaThreadSynchronize();
        CudaCheckError();
        CudaSafeCall(cudaMemcpy(output,D_output_x,height*widthStep,cudaMemcpyDeviceToHost));
        cudaFree(D_output_x);
        cudaFree(garbage);
        cudaFreeArray(cuArray);
    }
    
    #include<cuda.h>
    #include<iostream>
    #include "CudaKernel.h"
    
    using namespace std;
    
    #define CudaSafeCall( err ) __cudaSafeCall( err, __FILE__, __LINE__ )
    #define CudaCheckError()    __cudaCheckError( __FILE__, __LINE__ )
    #define checkCudaErrors(err) __checkCudaErrors (err, __FILE__, __LINE__)
    
    
    texture <float,2,cudaReadModeElementType> tex1;
    
    static cudaArray *cuArray = NULL;
    
    //Kernel for x direction sobel
    __global__ void implement_x_sobel(float* output,int width,int height,int widthStep)
    {
        int x = blockIdx.x * blockDim.x + threadIdx.x;
        int y = blockIdx.y * blockDim.y + threadIdx.y;
    
        //Make sure that thread is inside image bounds
        if(x<width && y<height)
        {
            float output_value = (-1*tex2D(tex1,x-1,y-1)) + (0*tex2D(tex1,x,y-1)) + (1*tex2D(tex1,x+1,y-1))
                               + (-2*tex2D(tex1,x-1,y))   + (0*tex2D(tex1,x,y))   + (2*tex2D(tex1,x+1,y))
                               + (-1*tex2D(tex1,x-1,y+1)) + (0*tex2D(tex1,x,y+1)) + (1*tex2D(tex1,x+1,y+1));
    
            output[y*widthStep+x]=output_value;
        }
    
    }
    
    
    inline void __checkCudaErrors( cudaError err, const char *file, const int line )
    {
        if( cudaSuccess != err) {
            fprintf(stderr, "%s(%i) : CUDA Runtime API error %d: %s.\n",
                file, line, (int)err, cudaGetErrorString( err ) );
            exit(-1);
        }
    }   
    
    //Host Code
    inline void __cudaSafeCall( cudaError err, const char *file, const int line )
    {
    #ifdef CUDA_ERROR_CHECK
        if ( cudaSuccess != err )
        {
            printf("cudaSafeCall() failed at %s:%i : %s\n",
                file, line, cudaGetErrorString( err ) );
            exit( -1 );
        }    
    #endif
    
        return;
    }
    inline void __cudaCheckError( const char *file, const int line )
    {
    #ifdef CUDA_ERROR_CHECK
        cudaError err = cudaGetLastError();
        if ( cudaSuccess != err )
        {
            printf("cudaCheckError() failed at %s:%i : %s\n",
                file, line, cudaGetErrorString( err ) );
            exit( -1 );
        }
    #endif
    
        return;
    }
    
    void kernelcall(float* input,float* output,int width,int height,int widthStep)
    {
        cudaChannelFormatDesc channelDesc = cudaCreateChannelDesc<float>();
    
        CudaSafeCall(cudaMallocArray(&cuArray,&channelDesc,width,height));
    
        //Never use 1D memory copy if host and device pointers have different widthStep.
        // You don't know the width step of CUDA array, so its better to use cudaMemcpy2D...
        cudaMemcpy2DToArray(cuArray,0,0,input,widthStep,width * sizeof(float),height,cudaMemcpyHostToDevice);
    
        cudaBindTextureToArray(tex1,cuArray,channelDesc);
    
        float * D_output_x;
        CudaSafeCall(cudaMalloc(&D_output_x,widthStep*height)); 
    
        dim3 blocksize(16,16);
        dim3 gridsize;
        gridsize.x=(width+blocksize.x-1)/blocksize.x;
        gridsize.y=(height+blocksize.y-1)/blocksize.y;
    
        implement_x_sobel<<<gridsize,blocksize>>>(D_output_x,width,height,widthStep/sizeof(float));
    
        cudaThreadSynchronize();
        CudaCheckError();
    
        //Don't forget to unbind the texture
        cudaUnbindTexture(tex1);
    
        CudaSafeCall(cudaMemcpy(output,D_output_x,height*widthStep,cudaMemcpyDeviceToHost));
    
        cudaFree(D_output_x);
        cudaFreeArray(cuArray);
    }
    
    #包括
    #包括
    #包括“CudaKernel.h”
    使用名称空间std;
    #定义CudaSafeCall(err)\ CudaSafeCall(err、\文件、\行)
    #定义CudaCheckError()\uuuu CudaCheckError(\uuuuu文件\uuuuu,\uuuuu行\uuuu)
    #定义校验错误(err)\校验错误(err、\文件、\行)
    纹理tex1;
    静态cudaArray*cuArray=NULL;
    //x方向sobel核
    __全局\uuuuvoid工具\ux\uSobel(浮点*输出、整数宽度、整数高度、整数宽度步长)
    {
    intx=blockIdx.x*blockDim.x+threadIdx.x;
    int y=blockIdx.y*blockDim.y+threadIdx.y;
    //确保线程在图像边界内
    
    if(xok)我更改了代码,但现在我得到一个未处理的异常错误。处理后进行规格化。使用cvNormalize(输入,输入规格化,0255,NORM_MINMAX,CV_8UC1)因此,在Cuda内核调用后对其进行规范化?cvNormalize函数总是导致未经处理的异常错误,请查看我的帖子,我更新了代码您不能将normalize用于float*。必须进行规范化。通常,建议您就某个问题开始悬赏,而不是询问可能的重复项。如果您只是在搜索gpu sobel筛选器,gpu opencv已经提供了几个过滤函数,看起来它们也有一个。你的代码工作得很完美,也很简单,谢谢你指出我的错误!我真的很感激你的代码非常有用。但是我有一个问题:为什么要使用浮点数组来存储输入和输出?我可以使用无符号字符数组吗?我使用浮点,因为OP是有意的使用它。是的,您可以使用
    无符号字符
    。在这种情况下,您必须创建深度为
    IPL\U depth\U 8U
    的主机映像
    image2
    image3