CUDA精进之路(三):图像处理——图像灰度化、灰度直方图统计

版权声明:最终版权归YBAidam所有 https://blog.csdn.net/Aidam_Bo/article/details/89146505

引言


在大部分的图像处理程序中,其中必不可少的一步就是对传入的彩图进行灰度处理,将三个通道的RGB图片转化为单通道的Gray图,而对于灰度图进行直方图统计同样是观察检测图像特征的常用方法。在OpenCV中已经有成熟的封装函数进行上述功能的实现,本文主要讲述CUDA实现,加快对大图像的处理速度。

任务要求


输入一张彩色图片,通过CUDA将其转换为灰度图,并对灰度图进行灰度直方图统计。

实现思路


关于彩色图片灰度化原理可参照此博客图像灰度化总结简单来说就是将传入的RGB图中三个通道的像素值分别乘以一定系数即可得到对应像素点灰度值。而在灰度值统计时,由于在CUDA中block内的线程同时并行,故无法采用简单的累加统计,需用到CUDA内自带的原子操作函数atomicAdd()

实现代码

#include "cuda_runtime.h"
#include "device_launch_parameters.h"
#include <cuda.h>
#include <device_functions.h>
#include <opencv2\opencv.hpp>
#include <iostream>
#include <math.h>
using namespace std;
using namespace cv;

//输入图像为BGR图,将其转化为gray图
__global__ void rgb2grayInCuda(uchar3 *dataIn, unsigned char *dataOut, int imgHeight, int imgWidth)
{
    int xIndex = threadIdx.x + blockIdx.x * blockDim.x;
    int yIndex = threadIdx.y + blockIdx.y * blockDim.y;

    if (xIndex < imgWidth && yIndex < imgHeight)
    {
        uchar3 rgb = dataIn[yIndex * imgWidth + xIndex];

        dataOut[yIndex * imgWidth + xIndex] = 0.299f * rgb.x + 0.587f * rgb.y + 0.114f * rgb.z;
    }
}

//灰度直方图统计
__global__ void imHistInCuda(unsigned char *dataIn, int *hist)
{
    int threadIndex = threadIdx.x + threadIdx.y * blockDim.x;
    int blockIndex = blockIdx.x + blockIdx.y * gridDim.x;
    int index = threadIndex + blockIndex * blockDim.x * blockDim.y;

    atomicAdd(&hist[dataIn[index]], 1);

}

int main()
{
    //传入图片
    Mat srcImg = imread("out1.bmp");

    int imgHeight = srcImg.rows;
    int imgWidth = srcImg.cols;
    Mat grayImg(imgHeight, imgWidth, CV_8UC1, Scalar(0));//输出灰度图
    int hist[256];//灰度直方图统计数组
    memset(hist, 0, 256 * sizeof(int));

    //在GPU中开辟输入输出空间
    uchar3 *d_in;
    unsigned char *d_out;
    int *d_hist;

    cudaMalloc((void**)&d_in, imgHeight * imgWidth * sizeof(uchar3));
    cudaMalloc((void**)&d_out, imgHeight * imgWidth * sizeof(unsigned char));
    cudaMalloc((void**)&d_hist, 256 * sizeof(int));

    //将图像数据传入GPU中
    cudaMemcpy(d_in, srcImg.data, imgHeight * imgWidth * sizeof(uchar3), cudaMemcpyHostToDevice);
    cudaMemcpy(d_hist, hist, 256 * sizeof(int), cudaMemcpyHostToDevice);

    dim3 threadsPerBlock(32, 32);
    dim3 blocksPerGrid((imgWidth + threadsPerBlock.x - 1) / threadsPerBlock.x, (imgHeight + threadsPerBlock.y - 1) / threadsPerBlock.y);

    //灰度化
    rgb2grayInCuda << <blocksPerGrid, threadsPerBlock >> >(d_in, d_out, imgHeight, imgWidth);

    //灰度直方图统计
    imHistInCuda << <blocksPerGrid, threadsPerBlock >> >(d_out, d_hist);

    //将数据从GPU传回CPU
    cudaMemcpy(hist, d_hist, 256 * sizeof(int), cudaMemcpyDeviceToHost);
    cudaMemcpy(grayImg.data, d_out, imgHeight * imgWidth * sizeof(unsigned char), cudaMemcpyDeviceToHost);


    cudaFree(d_in);
    cudaFree(d_out);
    cudaFree(d_hist);

    return 0;
}


实现结果


输入图片 

img.jpg

输出图片 

output.jpg

通过比对发现CUDA输出结果与OpenCV输出结果一致~
 

猜你喜欢

转载自blog.csdn.net/Aidam_Bo/article/details/89146505