Subspace Enhancement and Colorization Network for Infrared Video Action Recognition

Human action recognition is an essential area of research in the field of computer vision. However, existing methods ignore the essence of infrared image spectral imaging. Compared with the visible modality with all three channels, the infrared modality with approximate single-channel pays more attention to the lightness contrast and loses the channel information. Therefore, we explore channel duplication and tend to investigate more appropriate feature presentations. We propose a subspace enhancement and colorization network (S\(^2\)ECNet) to recognize infrared video action recognition. Specifically, we apply the subspace enhancement (S\(^2\)E) module to promote edge contour extraction with subspace. Meanwhile, a subspace colorization (S\(^2\)C) module is utilized for better completing missing semantic information. What is more, the optical flow provides effective supplements for temporal information. Experiments conducted on the infrared action recognition dataset InfAR demonstrates the competitiveness of the proposed method compared with the state-of-the-arts.