作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2023, Vol. 49 ›› Issue (1): 234-241,249. doi: 10.19678/j.issn.1000-3428.0063909

• 图形图像处理 • 上一篇    下一篇

面向同步辐射光源图像的可并行智能压缩方法

符世园1,2, 张敏行1,2, 高宇1,2, 汪璐1,2, 程耀东1,2,3   

  1. 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;
    3. 中国科学院高能物理研究所 天府宇宙线研究中心, 成都 610041
  • 收稿日期:2022-02-12 修回日期:2022-03-15 发布日期:2022-07-05
  • 作者简介:符世园(1995-),女,博士研究生,主研方向为分布式存储、数据压缩、深度学习;张敏行,博士研究生;高宇,硕士研究生;汪璐,副研究员、博士;程耀东,研究员、博士。
  • 基金资助:
    国家自然科学基金“面向多数据中心的LHAASO科学大数据管理系统及关键技术研究”(12075268)。

Parallel Intelligent Compression Method for Synchrotron Radiation Light Source Images

FU Shiyuan1,2, ZHANG Minxing1,2, GAO Yu1,2, WANG Lu1,2, CHENG Yaodong1,2,3   

  1. 1. Institute of High Energy Physics, Chinese Academy of Sciences, Beijing 100049, China;
    2. University of Chinese Academy of Sciences, Beijing 100049, China;
    3. Tianfu Cosmic Ray Research Center, Institute of High Energy Physics, Chinese Academy of Sciences, Chengdu 610041, China
  • Received:2022-02-12 Revised:2022-03-15 Published:2022-07-05

摘要: 在建的高能同步辐射光源预计会产生海量原始数据,其中硬X射线实验线站产生的图像数据占比最高且具有高分辨率和高帧率的特点,亟需有效的无损压缩方法缓解存储和传输压力,然而现有通用无损压缩方法对该类图像压缩效果不佳,基于深度学习的无损压缩方法又耗时较长。结合同步辐射光源图像的特点,提出一种在保证图像压缩比前提下的可并行智能无损图像压缩方法。通过参数自适应的可逆分区量化方法,大幅缩小图像经过时间差分后的像素值分布范围,能够节省20%以上的存储空间。将以CNN为基础架构的时空学习网络C-Zip作为概率预测器,同时以数据集为单位过拟合训练模型进一步优化图像压缩比。针对压缩过程中耗时较长的算术编码过程,利用概率距离量化代替算术编码,结合深度学习进行无损编码,增加编码过程的并行度。实验结果表明,该方法的图像压缩比相比于PNG、FLIF等传统图像无损压缩方法提升了0.23~0.58,对于同步辐射光源图像具有更好的压缩效果。

关键词: 无损压缩, 同步辐射光源图像, 深度学习, 可逆量化, 概率距离

Abstract: The High-Energy Photon Source(HEPS) under construction generates a large amount of raw data.The images generated by the hard X-ray imaging beamline, which is characterized by a high resolution and high frame rate, occupy the largest storage space, and an effective lossless compression method is urgently needed to relieve storage and transmission.However, the general lossless compression methods cannot work satisfactorily for these types of images, and methods based on deep learning will cost significant time.In this study, we propose a lossless compression method for synchrotron radiation source images, fully utilizing the image characteristics.First, we propose a method for parameter adaptive reversible partition quantization to significantly narrow the pixel value distribution range of the images after temporal difference, saving the storage space by more than 20%.Second, a spatial-temporal learning network named C-Zip based on the Convolutional Neural Network(CNN) is proposed as the probability predictor.Different models are trained by overfitting for different datasets to further optimize the Compression Ratio(CR). Probability distance quantization is used to replace arithmetic coding as a lossless coding method combined with deep learning to increase the parallelism of the coding process and improve the time-consuming arithmetic coding during the compression process.The experimental results show that compared with general lossless compression methods, such as Portable Network Graphics(PNG) and Free Lossless Image Format(FLIF), the intelligent compression method for synchrotron radiation source images can be further improved by 0.23-0.58 in the CR and has a better compression effect.

Key words: lossless compression, synchrotron radiation light source image, deep learning, reversible quantification, probability distance

中图分类号: