C++ opencv ncc归一化互相关 使用SSE优化 具体实现代码

以下是使用 SSE 优化的归一化互相关（NCC）的 C++ 代码实现：

#include <iostream>
#include <opencv2/opencv.hpp>
#include <emmintrin.h>

using namespace cv;

void ncc_sse(const Mat& img1, const Mat& img2, Mat& result)
{
    int width = img1.cols - img2.cols + 1;
    int height = img1.rows - img2.rows + 1;
    result.create(height, width, CV_32FC1);

    __m128i zero = _mm_setzero_si128();
    __m128 sum1, sum2, sum3, sum4, mean1, mean2, std1, std2, ncc, temp;
    __m128i p1, p2, mask;
    float* ptr1, * ptr2, * res;
    int i, j, k;

    for (i = 0; i < height; i++)
    {
        ptr1 = (float*)img1.ptr(i);
        for (j = 0; j < width; j++)
        {
            sum1 = sum2 = sum3 = sum4 = _mm_set_ps1(0.f);
            for (k = 0; k < img2.rows; k++)
            {
                ptr2 = (float*)img2.ptr(k);
                p1 = _mm_loadu_si128((__m128i*)(ptr1 + j));
                p2 = _mm_loadu_si128((__m128i*)(ptr2));

                sum1 = _mm_add_ps(sum1, _mm_mul_ps(_mm_loadu_ps(ptr1 + j), _mm_loadu_ps(ptr2)));
                sum2 = _mm_add_ps(sum2, _mm_mul_ps(_mm_loadu_ps(ptr1 + j + 4), _mm_loadu_ps(ptr2 + 4)));
                sum3 = _mm_add_ps(sum3, _mm_mul_ps(_mm_loadu_ps(ptr1 + j + 8), _mm_loadu_ps(ptr2 + 8)));
                sum4 = _mm_add_ps(sum4, _mm_mul_ps(_mm_loadu_ps(ptr1 + j + 12), _mm_loadu_ps(ptr2 + 12)));
                ptr1 += img1.cols;
                ptr2 += img2.cols;
            }

            mean1 = _mm_set_ps1(cv::mean(img1(cv::Rect(j, i, img2.cols, img2.rows)))[0]);
            mean2 = _mm_set_ps1(cv::mean(img2)[0]);

            temp = _mm_mul_ps(_mm_sub_ps(sum1, _mm_mul_ps(mean1, mean2)), _mm_rcp_ps(_mm_set_ps1(img2.cols * img2.rows)));
            sum2 = _mm_mul_ps(_mm_sub_ps(sum2, _mm_mul_ps(mean1, mean2)), _mm_rcp_ps(_mm_set_ps1(img2.cols * img2.rows)));
            sum3 = _mm_mul_ps(_mm_sub_ps(sum3, _mm_mul_ps(mean1, mean2)), _mm_rcp_ps(_mm_set_ps1(img2.cols * img2.rows)));
            sum4 = _mm_mul_ps(_mm_sub_ps(sum4, _mm_mul_ps(mean1, mean2)), _mm_rcp_ps(_mm_set_ps1(img2.cols * img2.rows)));

            std1 = _mm_sqrt_ps(_mm_sub_ps(_mm_mul_ps(temp, temp), _mm_mul_ps(mean1, mean1)));
            std2 = _mm_sqrt_ps(_mm_sub_ps(_mm_mul_ps(sum2, sum2), _mm_mul_ps(mean2, mean2)));
            std1 = _mm_mul_ps(_mm_rcp_ps(std1), _mm_rcp_ps(std2));

            ncc = _mm_mul_ps(temp, std1);
            res = (float*)result.ptr(i);
            _mm_storeu_ps(res + j, ncc);
        }
    }
}

int main()
{
    Mat img1 = imread("img1.png", IMREAD_GRAYSCALE);
    Mat img2 = imread("img2.png", IMREAD_GRAYSCALE);

    Mat result;
    ncc_sse(img1, img2, result);

    imshow("Result", result);
    waitKey(0);

    return 0;
}

其中，`_mm_set_ps1` 函数用于将一个值复制到 SSE 寄存器的四个数据位置；`_mm_loadu_si128` 函数用于从内存中加载一个 __m128i 类型的值；`_mm_add_ps` 函数用于两个 SSE 浮点数寄存器中的数据相加；`_mm_mul_ps` 函数用于两个 SSE 浮点数寄存器中的数据相乘；`_mm_set_ps` 函数用于将四个单精度浮点数组成一个 SSE 浮点数寄存器；`_mm_storeu_ps` 函数用于将一个 SSE 浮点数寄存器的值存储到内存中。