Learned Image Compression with Gaussian-Laplacian-Logistic Mixture Model and Concatenated Residual Modules

TL;DR

This work tackles the bottleneck of fixed, single-distribution entropy models in learned image compression by introducing a discretized Gaussian-Laplacian-Logistic Mixture Model (GLLMM) that jointly models latents with Gaussian, Laplacian, and Logistic components. Paired with a Concatenated Residual Module (CRM) to bolster the encoder’s learning capacity, the approach yields state-of-the-art RD performance on standard datasets, surpassing traditional codecs like VVC in PSNR and MS-SSIM. The method leverages a hyperprior and an autoregressive context model, optimized via a rate–distortion objective L = $\lambda D(\bm{x},\hat{\bm{x}}) + H(\hat{\bm{y}}) + H(\hat{\bm{z}})$, and demonstrates robust gains across Kodak, Tecnick-100, and Tecnick-40 with ablations highlighting the contributions of GLLMM and CRM. While the autoregressive context improves compression, it introduces serial decoding latency, pointing to future work on reducing complexity and exploring low-cost multivariate models for broader impact.

Abstract

Recently deep learning-based image compression methods have achieved significant achievements and gradually outperformed traditional approaches including the latest standard Versatile Video Coding (VVC) in both PSNR and MS-SSIM metrics. Two key components of learned image compression are the entropy model of the latent representations and the encoding/decoding network architectures. Various models have been proposed, such as autoregressive, softmax, logistic mixture, Gaussian mixture, and Laplacian. Existing schemes only use one of these models. However, due to the vast diversity of images, it is not optimal to use one model for all images, even different regions within one image. In this paper, we propose a more flexible discretized Gaussian-Laplacian-Logistic mixture model (GLLMM) for the latent representations, which can adapt to different contents in different images and different regions of one image more accurately and efficiently, given the same complexity. Besides, in the encoding/decoding network design part, we propose a concatenated residual blocks (CRB), where multiple residual blocks are serially connected with additional shortcut connections. The CRB can improve the learning ability of the network, which can further improve the compression performance. Experimental results using the Kodak, Tecnick-100 and Tecnick-40 datasets show that the proposed scheme outperforms all the leading learning-based methods and existing compression standards including VVC intra coding (4:4:4 and 4:2:0) in terms of the PSNR and MS-SSIM. The source code is available at \url{https://github.com/fengyurenpingsheng}

Figures (17)

• Figure 1: The block diagram of a typical image compression system.
• Figure 2: The Framework of the proposed image compression scheme. $G$ and $IG$ represent the GDN module and inverse GDN module, respectively. $\uparrow$ and $\downarrow$ denote the up- or down- sampling. $3 \times 3$ is the size of convolution kernel. $Q$ represents quantization. $AE$ and $AD$ represent the arithmetic encoder and arithmetic decoder, respectively. $L$ represents leaky ReLU. The dotted lines denote the shortcut connection with size change, as in resblockcheng2020.
• Figure 3: (a) The standard residual block; (b) the residual block with Leaky ReLU; (c) the proposed two-stage concatenated residual module; (d) the proposed three-stage concatenated residual module.
• Figure 4: Illustration of the proposed entropy coding model.
• Figure 5: The average PSNR and MS-SSIM performances of different methods on all 24 Kodak images.