Fast lossless images compression for synchrotron radiation facility using deep learning and hybrid architecture

The rapid growth in image data generated by high-energy photon sources poses significant challenges for storage and analysis, with conventional compression methods offering compression ratios often below 1.5.

This study introduces a novel, fast lossless compression method that combines deep learning with a hybrid computing architecture to overcome existing compression limitations. By employing a spatiotemporal learning network for predictive pixel value estimation and a residual quantization algorithm for efficient encoding.

When benchmarked against the DeepZip algorithm, our approach demonstrates a 40% reduction in compression time while maintaining comparable compression ratios using identical computational resources. The implementation of a GPU + CPU + FPGA hybrid architecture further accelerates compression, reducing time by an additional 38%.

This study presents an innovative solution for efficiently storing and managing large-scale image data from synchrotron radiation facilities, harnessing the power of deep learning and advanced computing architectures.