Context-dependent entropy for 3D hyperspectral image compression and reconstruction

Rapid advancements in hyperspectral (HS) methodologies for image analysis have resulted in specialized HS tasks, well-known for their extensive spatial-spectral data. Spectral bands provide the capability to distinguish between substance spectra, crucial for material analysis. However, the high-dimensional data volume of HS images (HSI) poses challenges for data storage. To address this issue, a new proximal process for context modeling is suggested, leveraging global similarity within the context. This approach effectively tackles compressive storage handling and context-based entropy modeling in large images. A full complex coder can provide minimally complicated encoding with a precise compression modeling scheme. The inherent spatial and spectral correlations, along with their subspaces, are effectively explored in these HS images. A key component of the performance is the entropy-based prediction of the probabilistic distribution code. Conventional deep learning (DL)-based entropy modeling techniques often overlook global similarity through the context, assuming that latent codes are systematically independent or dependent on extraneous data or local context. To address this, spatial masks and proximal splitting with similarity functions are employed to mitigate the issue of missing references. The compressed context-based entropy model (CCEM) is adapted to bridge the local and global context and applied to the concealed deep learning network (CDLN) to model entropy for better reconstruction progress compared to relevant studies.