When a standard lossy compression algorithm encounters a noisy signal, it faces a dilemma. Traditional codecs are designed under the assumption that every high-frequency variation in the input represents meaningful detail. Because they cannot distinguish between true underlying structure and random sensor noise, they suffer from what Ka Leong Cheng et al. term the bit misallocation problem in Optimizing Image Compression via Joint Learning with Denoising. The encoder wastefully expends its limited bit budget trying to faithfully reconstruct stochastic grain rather than preserving the core signal.
As highlighted by Jin, Jian et al. in Powerful Lossy Compression for Noisy Images and the authors of All-in-One Image Compression and Restoration, this behavior severely degrades compression efficiency. Paradoxically, in low-level machine vision pipelines, allocating more bits to a noisy source can actually harm downstream performance. As Qi Zhang 0042 et al. observe in LL-ICM: Image Compression for Low-level Machine Vision via Large Vision-Language Model, higher bitrates preserve more of the chaotic, high-frequency noise that interferes with generative restoration models.
However, this liability conceals a powerful mathematical symmetry. If we invert our perspective, the inherent difficulty of compressing noise can be leveraged as an unsupervised statistical filter. By forcing data through a tight information bottleneck, lossy compression naturally acts as a denoiser.
The Information-Theoretic Asymmetry
Why does compression possess an intrinsic capacity to denoise? The answer lies in the fundamental asymmetry between the entropy of structured signals and the entropy of random noise.
Natural signals—whether images, time series, or speech—possess underlying structural patterns, meaning they have low conditional entropy and are highly compressible. Random noise, by definition, is high-entropy and lacks predictable dependencies. In JPEG Information Regularized Deep Image Prior for Denoising, Tsukasa Takagi et al. demonstrated a predictable, monotonic relationship between an image's noise level and its compressed file size: because random noise resists standard compression algorithms, adding noise directly inflates the final bit count.
Noisy Input ───► [ Entropy Bottleneck ] ───► Clean Reconstruction
(Signal + Noise) (Strict Bit Budget) (Low Entropy Only)
│
└─► High-Entropy Noise Discarded
(Too expensive to encode)
From the perspective of the Minimum Description Length (MDL) principle, representing stochastic noise requires a disproportionately high description cost. In Minimum description length as an objective function for non-negative matrix factorization, Adam Prugel-Bennett et al. show that treating data decomposition as a compression task prevents a model from fitting noise. Because the description cost of random fluctuations is too steep, the optimization objective naturally prioritizes the underlying signal. Similarly, Aleena Siji et al. leverage this robustness in Seqret: Mining Rule Sets from Event Sequences, demonstrating that compression-driven algorithms can recover ground-truth patterns even in sequences containing up to 80% noise, simply because spurious patterns fail to yield efficient data descriptions.
The Bottleneck as an Unsupervised Regularizer
When we optimize a system using a rate-distortion objective, we minimize a combination of the bit budget (rate) and reconstruction error (distortion). Under a strict rate constraint, the compressor is forced to be selective. To achieve the lowest possible distortion per bit, it must allocate its budget to the components that yield the highest reduction in error. Structured patterns offer a high return on investment; high-entropy noise does not.
Shirin Jalali and Tsachy Weissman formalized this framework in Rate-Distortion via Markov Chain Monte Carlo, treating lossy compression as a denoising mechanism that constrains a noisy signal to a lower-entropy representation. This mathematical behavior allows compression to serve as a zero-shot, unsupervised regularizer for data restoration.
This principle underpins the work of Ali Zafari et al. in DeCompress: Denoising via Neural Compression and Zero-shot Denoising via Neural Compression: Theoretical and algorithmic framework. By training a neural compression model on patches of a single noisy image with an explicit entropy bottleneck, the network naturally learns to reconstruct the low-entropy signal while ignoring the expensive, stochastic noise. Because signal compressibility acts as a universal prior, this information-theoretic constraint generalizes across diverse noise types (such as Gaussian, Poisson, and real-world sensor noise) without requiring any clean training examples.
This filtering property manifests across multiple domains and data modalities:
- Pattern Alignment: In A Complexity-Distortion Approach to Joint Pattern Alignment, Andrea Vedaldi and Stefano Soatto frame image congealing as a lossy compression task, showing that the objective naturally filters out random pixel noise because it increases complexity without contributing to the shared structural pattern.
- Fluid Dynamics: Masahiro Takei et al. demonstrate in 102 PIV STANDARD IMAGE COMPRESSION BASED ON WAVELET TRANSFORM that discarding low-energy wavelet coefficients during compression functions as a non-linear denoiser, cleaning Particle Image Velocimetry data because random noise naturally concentrates in those low-magnitude coefficients.
- Biometrics: In Empirical Study of Image Compression for Palm Vein Recognition, Zhenhua Guo et al. show that light lossy compression acts as a low-pass filter that removes sensor artifacts, thereby improving downstream biometric recognition accuracy.
- Time Series Analytics: Matthias Boehm demonstrates in CAMEO: Autocorrelation-Preserving Line Simplification for Lossy Time Series Compression that compression algorithms designed to maintain statistical integrity (like autocorrelation) act as signal filters. By selectively removing points that deviate from the core temporal structure, they reduce noise and improve downstream forecasting. Similarly, Guoyou Sun and Qi Zhang 0020 show in Highly Efficient Direct Analytics on Semantic-aware Time Series Data Compression that lossy compression amplifies underlying semantic patterns, allowing outlier detection to achieve higher accuracy on compressed representations than on raw, noisy data.
Designing Noise-Aware Compression Pipelines
Understanding that compression naturally segregates data based on entropy has led to architectures that intentionally integrate denoising and compression into unified pipelines.
One prominent strategy is explicit signal separation prior to the bottleneck. In Compress Clean Signal from Noisy Raw Image: A Self-Supervised Approach, the authors introduce a self-supervised framework that uses a physical camera noise model to split raw images into a clean signal and a noise component before the compression bottleneck. By acting as a gatekeeper that prevents high-entropy noise from entering the bitstream, the system avoids wasting bandwidth on sensor grain, yielding massive bitrate savings and cleaner reconstructions.
┌───► Base Layer (Low Bitrate) ───► Clean Signal
Noisy Input ───┤ (Imposed Bitrate Constraint)
└───► Enhancement Layer ──────────► Stochastic Noise
(High Entropy Relegation)
Alternatively, this separation can be achieved structurally within the latent space itself. In Joint image compression and denoising via latent-space scalability, Saeed Ranjbar Alvar et al. introduce a scalable framework that splits image data into a base layer and an enhancement layer. By imposing strict bitrate constraints on the base latent layer, the network is incentivized to relegate high-entropy sensor noise to the secondary enhancement layer, exploiting the inherent compressibility gap to perform signal separation during encoding.
Ultimately, framing compression through the lens of optimal transport or posterior sampling—as explored by Berivan Isik and Tsachy Weissman in Learning under Storage and Privacy Constraints, Huan Liu et al. in Lossy Compression with Distribution Shift as Entropy Constrained Optimal Transport, and Nam Nguyen et al. in Cross-Domain Lossy Compression via Rate- and Classification-Constrained Optimal Transport—allows lossy compression to function as a statistical filter. When the distortion criteria of the compressor are properly aligned with the statistical properties of the noise, the information bottleneck naturally discards components that do not conform to the target data distribution. This transforms a resource constraint into a tool for signal enhancement.
Key Foundational Work
102 PIV STANDARD IMAGE COMPRESSION BASED ON WAVELET TRANSFORM (1999) — contributed 2 concept elements to this direction.
A Complexity-Distortion Approach to Joint Pattern Alignment (2006) — contributed 1 concept elements to this direction.