Quality Assessment of Deep-Learning-Based Image Compression

Abstract

Image compression standards rely on predictive coding, transform coding, quantization and entropy coding, in order to achieve high compression performance. Very recently, deep generative models have been used to optimize or replace some of these operations, with very promising results. However, so far no systematic and independent study of the coding performance of these algorithms has been carried out. In this paper, for the first time, we conduct a subjective evaluation of two recent deeplearning-based image compression algorithms, comparing them to JPEG 2000 and to the recent BPG image codec based on HEVC Intra. We found that compression approaches based on deep auto-encoders can achieve coding performance higher than JPEG 2000, and sometimes as good as BPG. We also show experimentally that the PSNR metric is to be avoided when evaluating the visual quality of deep-learning-based methods, as their artifacts have different characteristics from those of DCT or wavelet-based codecs. In particular, images compressed at low bitrate appear more natural than JPEG 2000 coded pictures, according to a no-reference naturalness measure. Our study indicates that deep generative models are likely to bring huge innovation into the video coding arena in the coming years.

Enhancing HEVC Spatial Prediction by Context-based Learning

“…Conversely, the CBPE can enhance this prediction, smoothing out the HEVC blocking and recovering somehow better the original image structure. Interestingly, the CBPE predictor has a more natural aspect, confirming previous findings on the ability of deep generative models to learn image "naturalness" [10].…”

“…Auto-encoder architectures [2,15], in particular, are especially effective to obtain compressed latent representations, by forcing the output to reproduce the input image through an information bottleneck whose dimensionality is much smaller than the original input space. Image compression methods based on auto-encoders have been shown to yield coding gains compared to legacy image codecs such as JPEG and JPEG 2000, and competitive results with more recent image compression algorithms such as BPG [7,12,10].…”

“…Differently from the signal models used in video spatial prediction, deep generative models are much more complex and highly non-linear, and can potentially approximate any class of signal provided that enough training samples are available to learn the original data distribution. In the last couple of years, these models have been applied to image compression [3,4,5,6,7,8,9], producing image representations able to provide, in some cases, equivalent or better visual quality than traditional image codecs [10].…”

Deep Medical Image Reconstruction with Autoencoders using Deep Boltzmann Machine Training

“…Valenzise et al [8] proposed a research article for image compression algorithm analysis based on subjective and objective parameters with deep learning-based image compression algorithms. Performance is analyzed in comparing with the JPEG and JPEG2000 lossless algorithms.…”

Perceptual Quality Study on Deep Learning Based Image Compression

“…Very few, however, have put emphasis on evaluating their performance based on subjective quality assessments. An example is the work in [17] which relies on small images (736 × 960) not necessarily fit for the current trend in highresolution imaging applications. A particularly important observation is that learned compression brings new types of artifacts that differ from blocking or ringing artifacts created by traditional codecs, as illustrated in Fig.…”