Fig. 5: Selected experimental results specifically aimed at the zero-shot generalization capability of the semantically regularized 3D compressive microwave meta-imaging.
From: Semantic regularization of electromagnetic inverse problems
a Reconstructions based on semantic embeddings including unseen language-based control commands to manipulate the reconstruction. \({{{{{\boldsymbol{{\ell}}}}}}}^{0}\) corresponds to the original semantic embedding \({{{{{{\boldsymbol{\alpha }}}}}}}_{0}^{0}\) proposed by the encoder network; \({{{{{\boldsymbol{{\ell}}}}}}}^{1}\)-\({{{{{\boldsymbol{{\ell}}}}}}}^{3}\) correspond to the changed semantic embeddings \({{{{{{\boldsymbol{\alpha }}}}}}}_{0}^{1}\)-\({{{{{{\boldsymbol{\alpha }}}}}}}_{0}^{3}\) which have similar meanings as \({{{{{{\boldsymbol{\alpha }}}}}}}_{0}^{0}\) but are not included in the training dataset; \({{{{{\boldsymbol{{\ell}}}}}}}^{4}\)-\({{{{{\boldsymbol{{\ell}}}}}}}^{6}\) corresponds to the changed semantic embeddings \({{{{{{\boldsymbol{\alpha }}}}}}}_{0}^{4}\)-\({{{{{{\boldsymbol{\alpha }}}}}}}_{0}^{6}\) which are semantically similar but different from \({{{{{{\boldsymbol{\alpha }}}}}}}_{0}^{0}\), and only \({{{{{{\boldsymbol{\alpha }}}}}}}_{0}^{4}\) is included in the training dataset. b T-SNE scatter plot visualizing the resemblance of the semantic embeddings considered in (a), \({{{{{\rm{i}}}}}}.{{{{{\rm{e}}}}}}.,{{{{{{\boldsymbol{\alpha }}}}}}}_{0}^{0}-{{{{{{\boldsymbol{\alpha }}}}}}}_{0}^{6}\).
