Transformer-Based Neural Video Compression on Solar Imagery

Abstract

NASA's Solar Dynamics Observatory (SDO) mission gathers extensive data on the Sun's daily activities. For space missions, data compression is essential to minimize data storage and video bandwidth needs by eliminating data redundancies. In this paper, we introduce an innovative neural Transformer-based approach for video compression, tailored specifically for SDO images. Our main goal is to efficiently leverage both temporal and spatial redundancies inherent in solar images to achieve a substantial compression ratio. Our proposed architecture incorporates a distinctive Transformer block termed Fused Local-aware Window (FLaWin). This block integrates window-based self-attention modules and an efficient Fused Local-aware Feed-Forward (FLaFF) network. This unique design allows us to simultaneously capture short-range and long-range information while facilitating the extraction of diverse and comprehensive contextual representations. Furthermore, this design choice results in a reduction of computational complexity. Experimental findings underscore the significant contribution of the FLaWin Transformer block to compression performance, surpassing conventional hand-engineered video codecs like H.264 and H.265 in terms of rate-distortion performance.