Abstract
Recent research has explored integrating lattice vector quantization (LVQ) into learned image compression models. Due to its more efficient Voronoi covering of vector space than scalar quantization (SQ), LVQ achieves better rate-distortion (R-D) performance than SQ, while still retaining the low complexity advantage of SQ. However, existing LVQ-based methods have two shortcomings: 1) lack of a multirate coding mode, hence incapable to operate at different rates; 2) the use of a fixed lattice basis, hence nonadaptive to changing source distributions. To overcome these shortcomings, we propose a novel adaptive LVQ method, which is the first among LVQ-based methods to achieve both rate and domain adaptations. By scaling the lattice basis vector, our method can adjust the density of lattice points to achieve various bit rate targets, achieving superior R-D performance to current SQ-based variable rate models. Additionally, by using a learned invertible linear transformation between two different input domains, we can reshape the predefined lattice cell to better represent the target domain, further improving the R-D performance. To our knowledge, this paper represents the first attempt to propose a unified solution for rate adaptation and domain adaptation through quantizer design.
Introduction
Most modern neural image compression systems rely on Scalar Quantization (SQ)—quantizing each feature independently—due to its implementation simplicity. However, information theory suggests that Vector Quantization (VQ)—quantizing groups of features together—is mathematically more efficient. Lattice VQ (LVQ) is a structured form of VQ that offers the efficiency of VQ with the low complexity of SQ.
Historically, LVQ models have been limited by their inflexibility. Changing the bitrate typically required training and storing entirely separate models, and fixed quantizers often struggled with "out-of-domain" data like screen captures or digital art. This paper introduces an Adaptive LVQ framework that provides a unified solution for both rate and domain adaptation through geometric manipulation of the lattice structure.
Methods: Adaptive Lattice Vector Quantization
The core innovation lies in defining the lattice basis $B$ as the product of a fixed "mother" lattice $G$ and a learnable, invertible linear transformation $A$ ($B = AG$).
Rate and Domain Adaptation
- Rate Adaptation: The transformation $A$ is implemented as a diagonal scaling matrix. By adjusting a single scaling factor, the model can "shrink" or "expand" the lattice cells. Smaller cells provide higher precision (higher bitrate, lower distortion), while larger cells provide higher compression. This allows a single network to cover a wide range of R-D trade-offs.
- Domain Adaptation: When transitioning to a new domain (e.g., cartoons), the main neural network remains frozen while a new transformation matrix $A$ is learned. This reshapes the lattice cells to better fit the specific feature distribution of the new domain.
Optimization
To handle the non-differentiable nature of quantization, the authors employ a Straight-Through Estimator (STE), ensuring that gradients can flow back through the quantizer during the training process.
Key Insights: The Voronoi Advantage
The superiority of LVQ over SQ is rooted in the geometry of its "Voronoi cells"—the regions of space assigned to a single lattice point. As shown in the figure above, while SQ uses simple cubic cells, LVQ can utilize shapes like hexagons or diamonds. These shapes can cover the feature space more efficiently, reducing the average quantization error for a given number of points.
Furthermore, the domain adaptation method is extremely lightweight. Learning the transformation matrix $A$ involves updating only about 0.06% of the total network parameters, providing a "near-free" performance boost on niche datasets without risking the "catastrophic forgetting" of the original natural image domain.
Experimental Results
The adaptive LVQ was integrated into the Cheng2020 architecture and evaluated against several benchmarks, including Flickr2W for training and Kodak, CLIC, and SCI1K for evaluation.
- Compression Performance: On the Tecnick dataset, the LVQ-based variable rate model achieved a BD-Rate reduction of 17.39% to 21.29% compared to standard SQ-based models. This indicates that the model can maintain image quality while using approximately 20% less data.
- Variable Rate Efficiency: The method significantly closes the performance gap between "variable-rate" (universal) models and "fixed-rate" (specialized) models, offering high performance with a much smaller memory footprint.
- Domain Adaptation Success: For out-of-domain tasks, adjusting only the quantizer improved performance on cartoon images by up to 1.59% BD-Rate.
Conclusion
By treating rate and domain adaptation as geometric transformations of a lattice, this work provides a flexible and efficient framework for neural image compression. While the gains in domain adaptation are currently modest, the unified approach to quantizer design represents a significant step toward more versatile and memory-efficient compression models. Future research will explore combining lattice optimization with selective fine-tuning of neural layers to further enhance performance on specialized image types.
Summary updated