C++版的LLC代码

图像稀疏编码总结：LLC和SCSPM，文章对稀疏编码讲解非常详细。

《Locality-constrained Linear Coding for Image Classification》的作者提供Matlab的代码实现，见http://www.ifp.illinois.edu/~jyang29/LLC.htm。

下面是根据作者的代码，基于OpenCV，实现的C++版的LLC：

Matlab Code：www.ifp.illinois.edu/~jyang29/LLC.htm

<span style="font-size:12px;">    cv::Mat BowModel::findKNN(cv::Mat &codebook, cv::Mat &input, int k) {
      int nbase = codebook.rows;
      int nquery = input.rows;
      Mat ii = input.mul(input);
      Mat cc = codebook.mul(codebook);  

      Mat sii(nquery,1,CV_32FC1);
      sii.setTo(0);
      Mat scc(nbase,1,CV_32FC1);
      scc.setTo(0);
      for (int i = 0; i<ii.rows; i++) {
        for (int j = 0; j<ii.cols; j++) {
            sii.at<float>(i,0) += ii.at<float>(i,j);
        }
      }
      for (int i = 0; i<cc.rows; i++) {
        for (int j = 0; j<cc.cols; j++) {
            scc.at<float>(i,0) += cc.at<float>(i,j);
        }
      }  

      Mat D(nquery,nbase,CV_32FC1);
      for (int i = 0; i<nquery; i++) {
        for (int j = 0; j<nbase; j++) {
            D.at<float>(i,j) = sii.at<float>(i,0);
        }
      }  

      Mat ct;
      transpose(codebook, ct);
      Mat D1 = 2*input*ct;  

      Mat scct;
      transpose(scc, scct);
      Mat D2(nquery, nbase, CV_32FC1);
      for (int i = 0; i<nquery; i++) {
        for (int j = 0; j<nbase; j++) {
            D2.at<float>(i,j) = scct.at<float>(0,j);
        }
      }  

      D = D - D1 + D2;
      Mat SD;
      sortIdx(D, SD, CV_SORT_EVERY_ROW+CV_SORT_ASCENDING);
      Mat IDX(nquery,k,CV_8UC1);
      for (int i = 0; i<nquery; i++) {
        for (int j = 0; j<k; j++) {
            IDX.at<uchar>(i,j) = SD.row(i).col(j).at<uchar>(0,0);
        }
      }  

      ii.release();
      cc.release();
      sii.release();
      scc.release();
      D.release();
      ct.release();
      D1.release();
      scct.release();
      D2.release();
      SD.release();
      return IDX;
    }  

    cv::Mat BowModel::llccode(cv::Mat &codebook, cv::Mat &input, cv::Mat IDX, int k)
    {
        int nquery = input.rows;
        int nbase = codebook.rows;
        int dim = codebook.cols;  

        Mat II = Mat::eye(k, k, CV_32FC1);
        Mat Coeff(nquery,nbase,CV_32FC1);
        Coeff.setTo(0);
        Mat z;
        Mat z1(k,dim,CV_32FC1);
        Mat z2(k,dim,CV_32FC1);
        Mat C;
        Mat un(k,1,CV_32FC1);
        un.setTo(1);
        Mat temp;
        Mat temp2;
        Mat w;
        Mat wt;  

        for (int n = 0; n<nquery; n++) {
            for (int i = 0; i<k; i++) {
                for (int j = 0; j<dim; j++) {
                    z1.at<float>(i,j) = codebook.at<float>(IDX.at<uchar>(n,i),j);
                    z2.at<float>(i,j) = input.at<float>(n,j);
                }
            }
            z = z1 - z2;
            transpose(z, temp);
            C = z*temp;
            C = C + II*(1e-4)*trace(C)[0];
            invert(C,temp2);
            w = temp2*un;
            float sum_w=0;
            for (int i = 0; i<k; i++) {
                sum_w += w.at<float>(i,0);
            }
            w = w/sum_w;
            transpose(w, wt);
            for (int i = 0; i<k; i++) {
                Coeff.at<float>(n,IDX.at<uchar>(n,i)) = wt.at<float>(0,i);
            }
        }  

        II.release();
        z.release();
        z1.release();
        z2.release();
        C.release();
        un.release();
        temp.release();
        temp2.release();
        w.release();
        wt.release();  

        return Coeff;
    }  </span>