HCF: Hierarchical Cascade Framework for Distributed Multi-Stage Image Compression

TL;DR

The paper addresses the challenge of distributed multi-stage image compression, where quality levels must adapt across networked processing nodes. It introduces the Hierarchical Cascade Framework (HCF), which uses direct latent-space transformations and policy-driven quantization to enable end-to-end cascaded compression with inter- and intra-node processing. A key contribution is the edge quantization principle, supported by differential-entropy analysis, that places quantization early in the cascade to improve rate–distortion performance by up to 0.6 dB PSNR and achieve substantial computational savings. Across multiple architectures and datasets, HCF outperforms state-of-the-art progressive and distributed baselines by up to 12.64% BD-Rate in PSNR, while delivering up to 97.8% FLOPs reduction, 96.5% GPU memory savings, and 90.0% faster execution, and enables retraining-free cross-quality adaptation with additional BD-Rate gains. These results establish a new paradigm for learned compression across transmission stages, with implications for scalable, efficient multimedia systems and potential extensions to video and adaptive policy control.

Abstract

Distributed multi-stage image compression -- where visual content traverses multiple processing nodes under varying quality requirements -- poses challenges. Progressive methods enable bitstream truncation but underutilize available compute resources; successive compression repeats costly pixel-domain operations and suffers cumulative quality loss and inefficiency; fixed-parameter models lack post-encoding flexibility. In this work, we developed the Hierarchical Cascade Framework (HCF) that achieves high rate-distortion performance and better computational efficiency through direct latent-space transformations across network nodes in distributed multi-stage image compression systems. Under HCF, we introduced policy-driven quantization control to optimize rate-distortion trade-offs, and established the edge quantization principle through differential entropy analysis. The configuration based on this principle demonstrates up to 0.6dB PSNR gains over other configurations. When comprehensively evaluated on the Kodak, CLIC, and CLIC2020-mobile datasets, HCF outperforms successive-compression methods by up to 5.56% BD-Rate in PSNR on CLIC, while saving up to 97.8% FLOPs, 96.5% GPU memory, and 90.0% execution time. It also outperforms state-of-the-art progressive compression methods by up to 12.64% BD-Rate on Kodak and enables retraining-free cross-quality adaptation with 7.13-10.87% BD-Rate reductions on CLIC2020-mobile.

Figures (6)

• Figure 1: The Distributed Recompression Framework (DRF) extends SIC to distributed compression across multiple nodes. It repeats the decompression-compression cycle at intermediate nodes. Specifically, each Node$_k$ compresses the image at quality level $s-k$ followed by the full decompression operations of quality level $s-k+1$.
• Figure 2: Hierarchical Cascade Framework (HCF). (a) Three $4$-level compression examples in a $2$-stage compression system, each corresponding to a different policy vector $\boldsymbol{\pi} \in \{[1,0,0,1], [0,1,0,1], [0,0,1,1]\}$ that determines the quantization placement in the transformation chain. At each level $k$, the process type follows the definition in Eq. \ref{['eq:selective_quant']}. (b) Architecture of transform modules $\phi_{k \rightarrow k-1}$ designed for cascade processing. $M$ denotes the number of channels in the latent representation. In convolutions, the change of the channel dimension is denoted by $\downarrow$ ($\times \frac{1}{2}$), $\uparrow$ ($\times 2$), and / (unchanged), respectively. The notation $M_k \rightarrow M_{k-1}$ indicates that the convolution maps the latent features from level $k$ to level $k-1$. All convolutions have stride $1$.
• Figure 3: Differential entropy evolution across compression levels for HCF's three quantization policies using MLIC++ on Kodak in a 2-stage, 4-level system. Stars indicate quantization points; $\uparrow$ show entropy increases after quantization.
• Figure 4: Visual quality comparison of HCF policies on Kodak dataset using MLIC++. $\boldsymbol{\pi}^{\text{edge}}=[1,0,0,1]$ preserves more details compared to alternatives at equivalent bitrates.
• Figure 5: Rate-distortion comparison on Kodak (PSNR) and CLIC (MS-SSIM) datasets. HCF with optimal $\boldsymbol{\pi}^{\text{edge}}$ policy consistently outperforms distributed baselines while approaching centralized SSF performance. Shaded areas indicate HCF's superior performance over DRF.