Algorithm 1: Self-Training Stage.

Input  : Set of labeled images: ${\mathcal{X}}^l$      Set of unlabeled images: ${\mathcal{X}}^u$      Net. init. weights & training hyp.-p.: $\mathcal{W},{\mathcal{H}}_{\mathcal{W}}$      Self-t. hyp.-p.: ${\mathcal{H}}_{st}=\{\mathcal{T},N,n,K_m,K_M,{\mathcal{M}}_{df}\}$ Output: Two refined models: ${\mathcal{W}}_{K_m},{\mathcal{W}}_{K_M}$                // Initialization              ${\mathcal{W}}_0$ ←$\mathrm{BasicModelTraining}(\mathcal{W},{\mathcal{H}}_{\mathcal{W}},{\mathcal{X}}^l)$  $<{\mathcal{X}}^{\widehat{l}},k,{\mathcal{V}}_{C_T},\mathcal{W}>$ ←$<\varnothing ,0,\mathbf{0},{\mathcal{W}}_0>$

return ${\mathcal{W}}_{K_m},{\mathcal{W}}_{K_M}$