Algorithm 1: Multiple-Criteria-based Patch Selection

Input:  Image patch set, $\mathcal{I}=\{I_1,I_2,\cdots ,I_m\}$  Number of textual patches, T  Number of cluster centers, k  Number of patches per cluster, n  Number of iterations, N  // Edge-and-Texture-based patch selection  for  $i=1,2,\dots ,m$  do   Calculate $f_i$ according to (2)  end for  Sort $f_i$ in descending order $f_{\pi \left(1\right)},f_{\pi \left(2\right)},\cdots ,f_{\pi \left(m\right)}$  Select the first T patches as edge and textual representatives: $\mathcal{E}=\{I_{\pi \left(1\right)},I_{\pi \left(2\right)},\cdots ,I_{\pi \left(T\right)}\})$  // Semantic-content-based patch selection  for  $i=1,2,\dots ,m$  do    ${\mu }_i=\frac{1}{3}{\sum }_{c\in [R,G,B]}{\mu }_{ci}$    ${\sigma }_i=\frac{1}{3}{\sum }_{c\in [R,G,B]}{\sigma }_{ci}$  end for  Form feature space $\mathbf{Z}=\{{\mathbf{\zeta }}_{\mathbf{1}},{\mathbf{\zeta }}_{\mathbf{2}},...,{\mathbf{\zeta }}_{\mathbf{m}}\})$ from patch set ${\mathbf{\zeta }}_{\mathbf{i}}=({\mu }_i,{\sigma }_i)$  Perform K-Means clustering in feature space $\mathbf{Z}$ to obtain the k Cluster centroids:  $c_1,c_2,\cdots ,c_k$ until N iterations is exceeded  For each of the k centroids, select n nearest patches as semantic representatives:  $\mathcal{S}=\{I_{c_{11}},I_{c_{12}},\cdots ,I_{c_{1n}},\cdots \cdots ,I_{c_{k1}},I_{c_{k2}},\cdots ,I_{c_{kn}}\}$    Output:  Training patch set $\mathcal{P}=\mathcal{E}\cup \mathcal{S}$