Home > News content

Compression efficiency first! CVPR image compression challenge Tencent audio video lab wins

via:博客园     time:2018/6/13 21:37:53     readed:355

The results of the CVPR 2018 image compression challenge (CLIC) have come out, and the Tencent audio and video laboratory and the joint team of Professor Chen Zhenzhong of Wuhan University have achieved compression performance first in the challenge.

CVPR is the top academic conference in the world. Since it was first held in 1983, it has a history of more than 30 years. It has a strong influence in the international academic circle. Every year, CVPR is a feast in the field of computer vision, and the top scholars, researchers and enterprises in the related fields of the world will actively participate. This year, CVPR added the image compression workshop and the challenge agenda, which was co sponsored by Google, Twitter, Amazon and other companies. It was the first image compression challenge launched by a conference on computer vision, which aims to introduce some new ways of neural network, deep learning, and so on into the field of image compression.

According to the official introduction of the conference, the challenge team evaluated the performance of the team from two aspects, PSNR and subjective evaluation respectively. The Tencent audio and video laboratory and the joint team iipTiramisu of Professor Chen Zhenzhong of Wuhan University took the lead in the PSNR (Peak Signal-to-Noise Ratio, peak signal to noise ratio) index, and were 0.13 higher than second in the final data set, ranking first. On other indicators, iipTiramisu is also in the forefront.

The peak signal to noise ratio (PSNR) is a measure of the error generated by two pixels by pixel statistics before and after compression. The higher the peak signal to noise ratio, the smaller the error of the two graph, which means that the closer the image to the original image and the smaller the loss of the quality of the picture.

IipTiramisu has a significant advantage over peak signal to noise ratio (SNR), compared with BPG, one of the best open source image compression algorithms in the industry.

orgsrc=https://images2018.cnblogs.com/news/66372/201806/66372-20180613212632234-1904681201.jpg

Fig.1 The Rate-PSNR curve of different encoders. (bit-rate range 0.05-0.35 BPP)

With constant signal-to-noise ratio, iipTiramisu saves 30.8% of the bit rate over BPG in performance-priority mode and 27.9% over BPG in speed-first mode.

orgsrc=https://images2018.cnblogs.com/news/66372/201806/66372-20180613212632035-175803564.jpg

orgsrc=https://images2018.cnblogs.com/news/66372/201806/66372-20180613212632290-1259235080.jpg

With a code rate of 0.14 bpp, the peak signal-to-noise ratio of iipTiramisu is 1.58 dB higher than that of BPG, and the quality loss rate is significantly reduced.

According to Professor Chen Zhenzhong, the iipTiramisu team used CNNMC and CNN in-loop filter coding based on hybrid image coder and deep learning, as well as uncertainty-based resource allocation strategy, which can ultimately compress dataset performance. This is a 30% increase over BPG.

iipTiramisu is formed by the Tencent audio and video laboratory Silicon Valley R&D center and the Wuhan University professor Chen Zhenzhong. Prof. Chen Zhenzhong is a professor and a professor at Wuhan University. He is a thousand-year-old person. He is mainly engaged in computer vision, image and video processing, human-computer interaction, and data mining. In recent years he has published more than 120 papers in international journals and has over 50 international projects. Proposals for domestic standards (H.265/HEVC/AVS), more than 10 international and domestic patent applications or authorizations. Tencent audio and video lab and Professor Chen Zhenzhong team have conducted in-depth cooperation in the fields of image video processing and artificial intelligence.

Image compression technology is of crucial importance to the transmission of information on the Internet. An uncompressed 12-megapixel image will occupy 36MB of storage space. At present, there are hundreds of millions of images transmitted and stored on the network every day. In order to save bandwidth resources and storage resources and reduce server pressure, it is highly efficient. Image compression algorithm is essential.

Tencent audio and video lab has a deep accumulation in the field of image compression. In May of last year, the laboratory launched a self-developed picture format TPG based on AVS. Its compression efficiency is also significantly ahead of JPG/JPEG, PNG, GIF, Mainstream image formats such as WEBP are at the world's leading level. In May of this year, TPG also won the AVS Industrial Technology Innovation Award from the AVS Working Group for its outstanding contributions in the development and promotion of AVS standards. This time, the compression efficiency of the CVPR 2018 Image Compression Challenge is the first, which means that Tencent Audio & Video Labs has made a great progress in image compression.

China IT News APP

Download China IT News APP

Please rate this news

The average score will be displayed after you score.

Post comment

Do not see clearly? Click for a new code.

User comments