Visual image analysis of high-rate local fixed-quality compression with neural networks

Fixed-quality compression is a lossy compression paradigm in which data is encoded at a user-defined tolerated distortion. When that tolerated distortion is allocated in certain regions of the image, it is referred to as local fixed-quality compression. It consists in compressing more complex regions at a higher bit rate and less complex areas at a lower bit rate, in order to reach the same level of quality with a reduced overall bit rate. Indeed, the global trend in high resolution satellite imagery is to drive the compression algorithm with a distortion target rather than with a bit rate target, which allows to allocate the bit rate resource where it is needed~\cite{camarero2015fixedquality}. The introduction of neural networks to data compression, particularly lossy image compression, has produced a breakthrough in the field. These novel methods outperform established algorithms that are the result of decades of innovation and fine-tuning. These advances have been applied to remote sensing satellite images through reduced-complexity algorithms that are compatible with on-board devices~\cite{alves2021mdpi, mijares2023multirate}. Indeed, missions such as ESA's Phi-Sat 2 will soon be pioneering the usage of neural compression on board remote sensing satellites~\cite{guerrisi2023phisat2}. Recent contributions have proposed methods for local fixed-quality compression using neural networks for remote sensing data, presenting rate-distortion and accuracy results for rates up to 1.5 bps or 2.5 bps~\cite{mijares2024fixedquality}. However in live applications target bit rates (and therefore image quality) are typically higher, so that distortion is at the level of instrumental noise. In this contribution, models for local fixed-quality compression at high qualities/bit rates are evaluated and visual image analysis is conducted on the artefacts. Furthermore, the computational capacity of the method is evaluated, showing it is compatible with new-generation on-board devices.