OpenCV学习指南(三)仿射变换

前言

  2-D下常见的矩阵变换如下图所示，本文主要介绍平移、旋转、放缩、剪切、反射，不包含透视变换。

  图像座标系和二维笛卡尔座标系之间存在一定差异，做仿射变换时，需要注意座标系转换带来的影响。
  上图是经典的二维笛卡尔座标系，下图图像座标系。

仿射变换：平移、旋转、放缩、剪切、反射

  以上各种仿射变换都是可以通过变换矩阵来实现的

  OpenCV进行各种变换用的函数是warpAffine

python中使用
cv2.warpAffine 
cpp的CPU版本
cv::warpAffine
Cuda版本使用
cv::cuda::warpAffine

  使用该函数需要注意四个参数，第一个输入的原图，第二个目标图，第三个，仿射变换矩阵(是个两行三列的矩阵)，第四个是目标图像的大小。

  笔者阅读了Cuda版本的仿射变换，发现，OpenCV的Cuda版本的仿射变换调用了CUDA的NPP库。而对于使用者而言，只要详细阅读相关文档，就能够很好的使用库。但对于开发者而言，先博而后渊，有了渊博的知识终能成为一个优秀的开发者。

使用实例

//cpp版本
//旋转angel弧度制+平移(x0,y0)
cv::Mat H = cv::Mat::zeros(2,3,CV_64FC1);
H.at<double>(0,0) = cos(angle);
H.at<double>(0,1) = -sin(angle);
H.at<double>(0,2) = x0;
H.at<double>(1,0) = sin(angle);
H.at<double>(1,1) = cos(angle);
H.at<double>(1,2) = y0;
cv::cuda::warpAffine(d_src,d_dst,H,_size);

作者说

  关于OpenCV的环境搭建，读者可以参见win10+ Opencv4+Cuda10+vs2017环境搭建；通过此链接可直接下载相关三方库。
  OpenCV源码分析系列的文章旨在介绍经典OpenCV库的同时介绍一定的图像处理知识，欢迎读者关注这个系列的文章。
  仅查阅资料可以通过OpenCV的官方文档

附录

OpenCV仿射变换的源代码

void cv::cuda::warpAffine(InputArray _src, OutputArray _dst, InputArray _M, Size dsize, int flags, int borderMode, Scalar borderValue, Stream& stream)
{
    GpuMat src = _src.getGpuMat();
    Mat M = _M.getMat();

    CV_Assert( M.rows == 2 && M.cols == 3 );

    const int interpolation = flags & INTER_MAX;

    CV_Assert( src.depth() <= CV_32F && src.channels() <= 4 );
    CV_Assert( interpolation == INTER_NEAREST || interpolation == INTER_LINEAR || interpolation == INTER_CUBIC );
    CV_Assert( borderMode == BORDER_REFLECT101 || borderMode == BORDER_REPLICATE || borderMode == BORDER_CONSTANT || borderMode == BORDER_REFLECT || borderMode == BORDER_WRAP );

    _dst.create(dsize, src.type());
    GpuMat dst = _dst.getGpuMat();

    Size wholeSize;
    Point ofs;
    src.locateROI(wholeSize, ofs);

    static const bool useNppTab[6][4][3] =
    {
        {
            {false, false, true},
            {false, false, false},
            {false, true, true},
            {false, false, false}
        },
        {
            {false, false, false},
            {false, false, false},
            {false, false, false},
            {false, false, false}
        },
        {
            {false, true, true},
            {false, false, false},
            {false, true, true},
            {false, false, false}
        },
        {
            {false, false, false},
            {false, false, false},
            {false, false, false},
            {false, false, false}
        },
        {
            {false, true, true},
            {false, false, false},
            {false, true, true},
            {false, false, true}
        },
        {
            {false, true, true},
            {false, false, false},
            {false, true, true},
            {false, false, true}
        }
    };

    bool useNpp = borderMode == BORDER_CONSTANT && ofs.x == 0 && ofs.y == 0 && useNppTab[src.depth()][src.channels() - 1][interpolation];
    // NPP bug on float data
    useNpp = useNpp && src.depth() != CV_32F;

    if (useNpp)
    {
        typedef void (*func_t)(const cv::cuda::GpuMat& src, cv::cuda::GpuMat& dst, double coeffs[][3], int flags, cudaStream_t stream);

        static const func_t funcs[2][6][4] =
        {
            {
                {NppWarp<CV_8U, nppiWarpAffine_8u_C1R>::call, 0, NppWarp<CV_8U, nppiWarpAffine_8u_C3R>::call, NppWarp<CV_8U, nppiWarpAffine_8u_C4R>::call},
                {0, 0, 0, 0},
                {NppWarp<CV_16U, nppiWarpAffine_16u_C1R>::call, 0, NppWarp<CV_16U, nppiWarpAffine_16u_C3R>::call, NppWarp<CV_16U, nppiWarpAffine_16u_C4R>::call},
                {0, 0, 0, 0},
                {NppWarp<CV_32S, nppiWarpAffine_32s_C1R>::call, 0, NppWarp<CV_32S, nppiWarpAffine_32s_C3R>::call, NppWarp<CV_32S, nppiWarpAffine_32s_C4R>::call},
                {NppWarp<CV_32F, nppiWarpAffine_32f_C1R>::call, 0, NppWarp<CV_32F, nppiWarpAffine_32f_C3R>::call, NppWarp<CV_32F, nppiWarpAffine_32f_C4R>::call}
            },
            {
                {NppWarp<CV_8U, nppiWarpAffineBack_8u_C1R>::call, 0, NppWarp<CV_8U, nppiWarpAffineBack_8u_C3R>::call, NppWarp<CV_8U, nppiWarpAffineBack_8u_C4R>::call},
                {0, 0, 0, 0},
                {NppWarp<CV_16U, nppiWarpAffineBack_16u_C1R>::call, 0, NppWarp<CV_16U, nppiWarpAffineBack_16u_C3R>::call, NppWarp<CV_16U, nppiWarpAffineBack_16u_C4R>::call},
                {0, 0, 0, 0},
                {NppWarp<CV_32S, nppiWarpAffineBack_32s_C1R>::call, 0, NppWarp<CV_32S, nppiWarpAffineBack_32s_C3R>::call, NppWarp<CV_32S, nppiWarpAffineBack_32s_C4R>::call},
                {NppWarp<CV_32F, nppiWarpAffineBack_32f_C1R>::call, 0, NppWarp<CV_32F, nppiWarpAffineBack_32f_C3R>::call, NppWarp<CV_32F, nppiWarpAffineBack_32f_C4R>::call}
            }
        };

        dst.setTo(borderValue, stream);

        double coeffs[2][3];
        Mat coeffsMat(2, 3, CV_64F, (void*)coeffs);
        M.convertTo(coeffsMat, coeffsMat.type());

        const func_t func = funcs[(flags & WARP_INVERSE_MAP) != 0][src.depth()][src.channels() - 1];
        CV_Assert(func != 0);

        func(src, dst, coeffs, interpolation, StreamAccessor::getStream(stream));
    }
    else
    {
        using namespace cv::cuda::device::imgproc;

        typedef void (*func_t)(PtrStepSzb src, PtrStepSzb srcWhole, int xoff, int yoff, float coeffs[2 * 3], PtrStepSzb dst, int interpolation,
            int borderMode, const float* borderValue, cudaStream_t stream, bool cc20);

        static const func_t funcs[6][4] =
        {
            {warpAffine_gpu<uchar>      , 0 /*warpAffine_gpu<uchar2>*/ , warpAffine_gpu<uchar3>     , warpAffine_gpu<uchar4>     },
            {0 /*warpAffine_gpu<schar>*/, 0 /*warpAffine_gpu<char2>*/  , 0 /*warpAffine_gpu<char3>*/, 0 /*warpAffine_gpu<char4>*/},
            {warpAffine_gpu<ushort>     , 0 /*warpAffine_gpu<ushort2>*/, warpAffine_gpu<ushort3>    , warpAffine_gpu<ushort4>    },
            {warpAffine_gpu<short>      , 0 /*warpAffine_gpu<short2>*/ , warpAffine_gpu<short3>     , warpAffine_gpu<short4>     },
            {0 /*warpAffine_gpu<int>*/  , 0 /*warpAffine_gpu<int2>*/   , 0 /*warpAffine_gpu<int3>*/ , 0 /*warpAffine_gpu<int4>*/ },
            {warpAffine_gpu<float>      , 0 /*warpAffine_gpu<float2>*/ , warpAffine_gpu<float3>     , warpAffine_gpu<float4>     }
        };

        const func_t func = funcs[src.depth()][src.channels() - 1];
        CV_Assert(func != 0);

        float coeffs[2 * 3];
        Mat coeffsMat(2, 3, CV_32F, (void*)coeffs);

        if (flags & WARP_INVERSE_MAP)
            M.convertTo(coeffsMat, coeffsMat.type());
        else
        {
            cv::Mat iM;
            invertAffineTransform(M, iM);
            iM.convertTo(coeffsMat, coeffsMat.type());
        }

        Scalar_<float> borderValueFloat;
        borderValueFloat = borderValue;

        func(src, PtrStepSzb(wholeSize.height, wholeSize.width, src.datastart, src.step), ofs.x, ofs.y, coeffs,
            dst, interpolation, borderMode, borderValueFloat.val, StreamAccessor::getStream(stream), deviceSupports(FEATURE_SET_COMPUTE_20));
    }
}