Deep Lossy Plus Residual Coding for Lossless and Near-lossless Image Compression

TL;DR

This work introduces Deep Lossy Plus Residual (DLPR) coding, a unified, end-to-end framework that enables both lossless and near-lossless image compression within a single neural architecture. By combining a high-capacity lossy image compressor with a residual compressor in a variational autoencoder (VAE) framework and adding autoregressive context for residuals, DLPR achieves state-of-the-art lossless performance while supporting scalable near-lossless with variable $\ell_\infty$ bounds. A scalable residual quantization strategy (SQRC) and a novel context design accelerate entropy coding, making the method practical for full-resolution images. Experimental results show that DLPR improves lossless and near-lossless RD performance across diverse datasets, while maintaining competitive runtimes compared with traditional and learned codecs, underscoring its potential for professional imaging tasks demanding high fidelity.

Abstract

Lossless and near-lossless image compression is of paramount importance to professional users in many technical fields, such as medicine, remote sensing, precision engineering and scientific research. But despite rapidly growing research interests in learning-based image compression, no published method offers both lossless and near-lossless modes. In this paper, we propose a unified and powerful deep lossy plus residual (DLPR) coding framework for both lossless and near-lossless image compression. In the lossless mode, the DLPR coding system first performs lossy compression and then lossless coding of residuals. We solve the joint lossy and residual compression problem in the approach of VAEs, and add autoregressive context modeling of the residuals to enhance lossless compression performance. In the near-lossless mode, we quantize the original residuals to satisfy a given $\ell_\infty$ error bound, and propose a scalable near-lossless compression scheme that works for variable $\ell_\infty$ bounds instead of training multiple networks. To expedite the DLPR coding, we increase the degree of algorithm parallelization by a novel design of coding context, and accelerate the entropy coding with adaptive residual interval. Experimental results demonstrate that the DLPR coding system achieves both the state-of-the-art lossless and near-lossless image compression performance with competitive coding speed.

Figures (17)

• Figure 1: Network architecture of DLPR coding framework, including a lossy image compressor (LIC), a residual compressor (RC) and a scalable quantized residual compressor (SQRC).
• Figure 2: Network architecture of lossy image compressor (LIC). We employ sophisticated image encoder/decoder while efficient hyper-prior model. The channel numbers are set uniformly to 192 for all layers. (AE: arithmetic encoding. AD: arithmetic decoding. Q: quantization.)
• Figure 3: Detailed structures of different blocks in LIC. (a) Analysis block. (b) Synthesis block. (c) Swin-Attention block. The window size and head number of Swin-Attention blocks are set to both $8$ for $4\times$ down-sampled feature maps, and are set to 4 and 8 for $16\times$ down-sampled feature maps. The channel numbers are set uniformly to 192 for all layers. (GDN: generalized divisive normalization gdn2016iclr. IGDN: inverse GDN. W-MSA/SW-MSA: window/shifted window based multi-head self-attention liu2021swin. ResBlock: residual block he2016deep.)
• Figure 4: Network architecture of entropy model in RC. Given $\mathbf u$ and $C_\mathbf r$, the entropy model estimates parameters of discrete logistic mixture likelihoods corresponding to the probability distribution of $\mathbf r$. All $1\times 1$ convolutional layers except the last layer have 256 channels. The last convolutional layer has $10\cdot K$ channels split by $\pi$, $\mu$, $\sigma$ and $\beta$.
• Figure 5: Probability inferences of residuals and quantized residuals. (a) Probability inference of residuals with RC. (b) Probability inference of quantized residuals with $\tau$-specific scheme. (c) Scalable probability inference of quantized residuals without SQRC.(d) Scalable probability inference of quantized residuals with SQRC. (RQ: residual quantization. PQ: PMF quantization.)

Theorems & Definitions (2)

• Proposition 1
• proof