“…Deep learning based methods are proposed in [103], [104] For matching visual and thermal face images, Chen et al [67] developed the high-frequency representation (HFR) framework, which matches images using multiple subspaces generated from patches. In the context of modality discrepancies, Ye et al [58] introduced the MACE learning method, which focuses on learning discriminative middle-level features and addresses the differences between modalities in features and classifiers [29], [58].…”