Multi‐scale spatial‐spectral attention network for multispectral image compression based on variational autoencoder

Abstract

Based upon the fact that multispectral image compression needs to remove both spatial and spectral redundancy, recent learnt models via end-to-end manners have shown promising performance. However, most of them ignore the characteristics of multispectral image, i.e., the non-stationarity of spectral correlation and the scale-diversity of spatial features. Meanwhile, they directly utilize fully factorized entropy model, rendering compression performance suboptimal. This paper proposes a Multi-Scale Spatial-Spectral Attention Network (MSSSA-Net) based on variational autoencoder (VAE). Our MSSSA-Net (1) incorporates a simple neuroscience-based non-local attention module into attention mechanism to capture the tiny features in adjacent pixels and large-scale features in spatial domain simultaneously, (2) proposes a multi-scale spectral attention block to extract non-stationary correlation of adjacent spectra at different scales. We demonstrate that our MSSSA-Net offers the state-of-the-art performance in comparison with classical algorithms, including JPEG2000 and 3D-SPIHT, and recent learnt image compression models, on 7-band and 8-band datasets from Landsat-8 and WorldView-3 satellites, when measured by PSNR, MS-SSIM and Mean Spectral Angle. Extensive ablation experiments have verified the effectiveness of each component, and have demonstrated that, for multispectral image compression, Scale-only Hyperprior can make a better trade-off between compression performance and complexity compared with Mean & Scale Hyperprior and Joint Autoregressive model.