Unified ROI-based Image Compression Paradigm with Generalized Gaussian Model

TL;DR

The paper tackles ROI-based image compression by addressing the mismatch between the sharp-peaked, heavy-tailed latent distributions and conventional Gaussian priors. It introduces a Generalized Gaussian Model (GGM) with learnable scale $\alpha$ and shape $\beta$, along with differentiable activations (Softplus for $\beta$, Hubber-like for $\alpha$) and a dynamic lower bound to stabilize training. A unified rate–distortion optimization paradigm is developed, integrating the GGM prior into RD objectives under spatial heterogeneity constraints and implicit bit allocation via a Mask-guided Feature Enhancement module. Empirical results on COCO2017 and HRSOD show state-of-the-art ROI reconstruction and improved performance on machine-vision tasks (segmentation and detection), with substantial BD-rate reductions and BD-PSNR gains compared to baselines. The approach achieves high fidelity in ROIs while maintaining efficiency, illustrating the practical impact of flexible distribution modeling and ROI-aware optimization for real-world vision systems.

Abstract

Region-of-Interest (ROI)-based image compression allocates bits unevenly according to the semantic importance of different regions. Such differentiated coding typically induces a sharp-peaked and heavy-tailed distribution. This distribution characteristic mathematically necessitates a probability model with adaptable shape parameters for accurate description. However, existing methods commonly use a Gaussian model to fit this distribution, resulting in a loss of coding performance. To systematically analyze the impact of this distribution on ROI coding, we develop a unified rate-distortion optimization theoretical paradigm. Building on this paradigm, we propose a novel Generalized Gaussian Model (GGM) to achieve flexible modeling of the latent variables distribution. To support stable optimization of GGM, we introduce effective differentiable functions and further propose a dynamic lower bound to alleviate train-test mismatch. Moreover, finite differences are introduced to solve the gradient computation after GGM fits the distribution. Experiments on COCO2017 demonstrate that our method achieves state-of-the-art in both ROI reconstruction and downstream tasks (e.g., Segmentation, Object Detection). Furthermore, compared to classical probability models, our GGM provides a more precise fit to feature distributions and achieves superior coding performance. The project page is at https://github.com/hukai-tju/ROIGGM.

Figures (18)

• Figure 1: (a) Comparison of latent feature histograms and fitted distributions, with corresponding Kullback-Leible (KL) divergence values. Our GGM can more accurately fit the peaked and heavy-tailed distribution of the latent variables, achieving the lowest KL divergence of 0.0224. (b) Achieved BD-rate savings compared to the GM baseline for each probability model. For both the entire image and the target region, our GGM achieves optimal coding performance in terms of BD-rate, significantly outperforming classical probability models.
• Figure 2: Analysis of the shape parameter $\beta$. (a) The effective gradient weight $\omega (\lvert e \rvert; \beta) = \beta \lvert e \rvert^{\beta - 1}$ for the distortion term $\mathcal{D}=\sum_i \omega_i \lvert e_i \rvert^{\beta}$. Setting $\beta=1$ gives constant weight; $\beta=2$ yields linear scaling; $\beta=3$ causes quadratic amplification of large errors. (b) Reconstruction error spectra on COCO2017. $\beta=1$ under-penalizes large errors, leading to heavy-tailed residuals; $\beta=3$ over-focuses on outliers; $\beta=2$ achieves the best balance. (c) Rate-distortion performance. The model with $\beta=2$ achieves superior reconstruction quality at lower bitrates.
• Figure 3: The influence of the scale ($\alpha$) and shape ($\beta$) parameters on the PDF of the GGM with a fixed mean ($\mu = 0$).
• Figure 4: The functions used for scale $\alpha$ and shape $\beta$. Lower bound values for shape $\beta$ and scale $\alpha$ need to be positive. It is known that Our functions are differentiable within the domain and have a wider effective input range.
• Figure 5: Feature maps with different functions for shape $\beta$ and scale $\alpha$.