Skip to main content

Research & Development

Development of image compression technology based on deep learning aiming for high perceptual image quality under an ultra-low-bitrate condition

Optimizing compression and decompression for each region in an image by importance of texture and detailed structure

September 30, 2020

  • Uncompressed


  • Compressed (JPEG)

    Compressed (JPEG)

  • Compressed (Ours)

    Compressed (Ours)

Fig. 1 Comparison of compressed images under similar bitrate conditions (384 x 384 pixels, click to see pop-up image)

Hitachi, Ltd. has developed an image compression technology optimized for each region in an image based on deep learning with the aim of efficiently utilizing vast quantities of image data that are generated daily. We aimed to optimize for a perceptual image quality rather than an index that simply expresses assessment through errors of pixel values.*1 This technology optimizes compression and decompression based on regions where texture is important or regions where detailed structure is important. It was validated that this technology realizes top-class perceptual image quality at an image compression competition*2 held under an ultra-low-bitrate condition*3 and judged by human evaluators. Hitachi will accelerate this research through co-creation, etc. and consider applying this method to IoT solutions utilizing vast quantities of data. In addition, Hitachi will also consider applying this method to support societal change under COVID-19 by researching needs for remote operation or automation, etc.


  • This work is part of joint research with Aizawa Laboratory of the Department of Information Communication Engineering, Graduate School of Information Science and Technology, the University of Tokyo.
  • Computational resources of AI Bridging Cloud Infrastructure (ABCI) provided by National Institute of Advanced Industrial Science and Technology (AIST) were used in this research.
Pixel value: a set of values indicating the color or brightness of a pixel, the smallest element from which images are made.
We participated in the Workshop and Challenge on Learned Image Compression (CLIC) 2020 at the Conference on Computer Vision and Pattern Recognition (CVPR) and won third place in the low-rate track.
A condition of no more than 0.15 bits per pixel on average for the entire test image dataset provided for CLIC 2020. This was a regulation for the competition.

For more information, use the enquiry form below to contact the Research & Development Group, Hitachi, Ltd. Please make sure to include the title of the article.

  • Page top