ARCHE: Autoregressive Residual Compression with Hyperprior and Excitation

Recent progress in learning-based image compression has demonstrated that end-to-end optimization can substantially outperform traditional codecs by jointly learning compact latent representations and probabilistic entropy models. However, many existing approaches achieve high rate-distortion efficiency at the expense of increased computational cost and limited parallelism. This paper presents ARCHE - Autoregressive Residual Compression with Hyperprior and Excitation, an end-to-end learned image compression framework that balances modeling accuracy and computational efficiency. The proposed architecture unifies hierarchical, spatial, and channel-based priors within a single probabilistic framework, capturing both global and local dependencies in the latent representation of the image, while employing adaptive feature recalibration and residual refinement to enhance latent representation quality. Without relying on recurrent or transformer-based components, ARCHE attains state-of-the-art rate-distortion efficiency: it reduces the BD-Rate by approximately 48% relative to the commonly used benchmark model of Balle et al., 30% relative to the channel-wise autoregressive model of Minnen & Singh and 5% against the VVC Intra codec on the Kodak benchmark dataset. The framework maintains computational efficiency with 95M parameters and 222ms running time per image. Visual comparisons confirm sharper textures and improved color fidelity, particularly at lower bit rates, demonstrating that accurate entropy modeling can be achieved through efficient convolutional designs suitable for practical deployment.

💡 Research Summary

The paper introduces ARCHE (Autoregressive Residual Compression with Hyperprior and Excitation), a novel end‑to‑end learned image compression framework that strives to balance high rate‑distortion performance with practical computational efficiency. Built upon a variational auto‑encoder backbone, ARCHE integrates four complementary modules: (1) a hierarchical hyper‑prior that transmits side‑information about the global statistics of the latent representation, enabling precise per‑element Gaussian parameter (mean μ and scale σ) estimation; (2) a masked PixelCNN‑style spatial autoregressive context model that conditions each latent code on previously decoded neighbours while respecting a causal mask, thereby preserving the benefits of autoregression without incurring the severe sequential bottleneck of full raster‑order decoding; (3) channel‑wise conditioning enhanced by a Squeeze‑and‑Excitation (SE) block, which learns dynamic scaling factors for each latent channel based on a global average‑pooled descriptor, correcting for channel‑specific non‑Gaussianity and multimodality; and (4) a lightweight residual‑prediction network that estimates quantization error and adds it back to the latent before synthesis, improving reconstruction especially at low bit‑rates.

All components are implemented with standard convolutions; no recurrent units or transformer blocks are used, keeping the architecture fully parallelizable on modern GPUs. The total parameter count is about 95 M, and decoding of a 1080p image on an RTX 3080 takes roughly 222 ms, a figure comparable to lightweight learned codecs and far faster than sequential context models.

Extensive experiments on the Kodak, Tecnick, and CLIC benchmarks demonstrate that ARCHE achieves state‑of‑the‑art compression quality: it reduces BD‑Rate by ~48 % relative to the classic Balle et al. hyper‑prior baseline, by 30 % compared with the channel‑wise autoregressive model of Minnen & Singh, and by 5 % against the VVC‑Intra intra‑frame codec. Visual comparisons highlight sharper textures and more faithful colors, particularly in the low‑bit‑rate regime where the residual‑prediction and SE modules are most effective.

The authors also provide a comparative table showing that ARCHE uniquely combines hyper‑prior, masked autoregressive context, channel conditioning, and excitation, whereas prior works typically adopt only a subset of these techniques. This holistic integration allows ARCHE to capture global, spatial, and channel dependencies simultaneously while preserving a high degree of parallelism.

In summary, ARCHE demonstrates that accurate entropy modeling does not require heavyweight transformer or recurrent architectures; a carefully designed convolutional pipeline that unifies hierarchical priors, masked context, channel recalibration, and residual correction can deliver both superior compression efficiency and deployment‑ready speed. The paper concludes with suggestions for future work, including more compact SE designs, multi‑scale hyper‑priors, and extensions to video compression.