基于深度学习的高能同步辐射光源数据快速无损压缩算法  

Deep learning-based fast lossless compression algorithm for data generated by High Energy Photon Source

作　　者：张敏行 符世园 高宇 程耀东[1,2] 陈刚 ZHANG Minxing;FU Shiyuan;GAO Yu;CHENG Yaodong;CHEN Gang(Institute of High Energy Physics,Chinese Academy of Sciences,Beijing 100049,China;University of Chinese Academy of Sciences,Beijing 100049,China)

机构地区：[1]中国科学院高能物理研究所,北京100049 [2]中国科学院大学,北京100049

出　　处：《计算机应用》2024年第S01期95-100,共6页journal of Computer Applications

基　　金：国家自然科学基金项目(12075268)。

摘　　要：高能同步辐射光源(HEPS)产生的高帧率、高分辨率图像数据以样本为单位存储,每个样本对应近千张灰度图。通用无损压缩方法对此类图像压缩效果不佳,而基于深度学习的压缩方法通常压缩解压耗时较长。为了解决上述问题,结合光源图像数据特点,提出一种基于深度学习的快速无损压缩算法。以样本对应的图像序列为单位压缩,根据少量前向帧利用时空学习网络预测图像取值并编码残差图,以实现高压缩比(CR);此外,通过图像子集划分和子集次优编码方法,将图像序列划分为多个子集并构建编码表,从而减少压缩解压缩耗时。经过测试,相较于PNG、JPEG2000等传统图像无损压缩算法,所提算法的CR提升了0.22~0.68;相较于DeepZip等深度学习无损压缩算法,所提算法的压缩耗时减少了41.0%以上,解压缩耗时减少了9.3%以上。The high frame rate,high-resolution image data generated by the High Energy Photon Source(HEPS)are stored on a per-sample basis,with each sample corresponding to nearly a thousand grayscale images.Lossless compression methods perform poorly on such images.Compression methods based on deep learning usually take a long time for compression and decompression.To address the above issues,a fast lossless compression algorithm based on deep learning was proposed combining the characteristics of the light source image data.The image sequences corresponding to samples were compressed by the proposed algorithm,image values were predicted based on a small number of preceding frames using a spatiotemporal learning network,and residual images were encodered to achieve a high Compression Ratio(CR).Furthermore,by partitioning the image sequences into multiple subsets and constructing encoding tables based on subset suboptimal coding,compression time and decompression time were reduced.Test results show that,compared to traditional lossless image compression algorithms such as PNG and JPEG2000,the proposed algorithm achieves a CR improvement of 0.22 to 0.68.Compared to the deep learning lossless compression algorithms such as DeepZip,the proposed algorithm reduces compression time by over 41.0%and decompression time by over 9.3%.

关 键 词：图像压缩 深度学习 图像分组 无损压缩 编码方法 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]