RAWIC: Adaptive Compression & RIS Sensing

Updated 5 July 2026

RAWIC is a dual-use framework that enables bit-depth adaptive lossless compression of Bayer raw sensor images, achieving significant bitrate reductions compared to standard codecs.
RAWIC also denotes a RIS-aided wireless imaging approach that integrates cooperative ISAC and sparse reconstruction to perform low-altitude surveillance under challenging conditions.
Both applications employ adaptive probabilistic modeling and structured measurement support to address the heterogeneous data and stringent sensing constraints inherent in their domains.

RAWIC denotes distinct but technically related concepts in recent arXiv literature. In learned image compression, RAWIC is the framework introduced in "RAWIC: Bit-Depth Adaptive Lossless Raw Image Compression," a camera-agnostic, bit-depth-adaptive learned lossless codec for Bayer-pattern raw sensor images that operates directly on raw measurements rather than on demosaiced or reconstructed proxies (Zheng et al., 30 Mar 2026). In RIS-aided sensing, RAWIC is used as a shorthand for reconfigurable intelligent surface aided wireless imaging/communications, exemplified by a cooperative ISAC formulation that treats low-altitude surveillance as a sparse 3D imaging problem solved from RIS-controlled channel measurements (Chen et al., 22 Jan 2026). A related theoretical backdrop is provided by restricted isometry random variable analysis, which replaces deterministic restricted isometry constants by ensemble-level random variables and yields sharper compressive-sensing predictions for Gaussian encoders (James et al., 2014).

1. Terminological scope

The term RAWIC is not monosemous in current technical usage. One usage belongs to raw image compression; another belongs to RIS-aided wireless imaging and communications. The overlap is methodological rather than semantic: both emphasize data-adaptive probabilistic modeling under structured sensing or measurement constraints.

Usage of RAWIC	Domain	Canonical formulation
RAWIC	Learned lossless raw image compression	Bit-depth-adaptive compression of Bayer raw images (Zheng et al., 30 Mar 2026)
RAWIC	RIS-aided wireless imaging/communications	Cooperative ISAC imaging for low-altitude surveillance (Chen et al., 22 Jan 2026)

This dual usage matters because the two literatures solve different problems. The compression framework addresses exact coding of raw sensor measurements with heterogeneous bit depths and camera statistics. The RIS-aided imaging formulation addresses low-altitude sensing under severe path loss by combining programmable propagation, compressed sensing, and cooperative ISAC. A plausible implication is that RAWIC functions more as a research label than as a single unified theory.

2. RAWIC as bit-depth-adaptive lossless raw image compression

In the compression literature, RAWIC is a learned lossless compression framework designed specifically for Bayer-pattern raw sensor images (Zheng et al., 30 Mar 2026). The task is exact coding of single-channel Bayer raw data from diverse cameras, where the data preserve linear radiance-related measurements, higher precision typically in the 10–14 bit range, and sensor-dependent statistics. The framework is motivated by the mismatch between raw data and standard learned lossless image compressors, which are usually built for fixed-format 8-bit per channel sRGB imagery.

The method addresses three stated difficulties: varying sensor bit depths across cameras, diversity across camera sensors, and intra-image variation in effective dynamic range. The paper motivates patch-wise adaptation using pixel-wise bit-depth maps computed as

$\left\lceil \log_2(\text{value}+1) \right\rceil,$

arguing that a fixed global bit depth wastes code space when local valid ranges are smaller. The framework therefore computes a bit depth for each patch and uses it as auxiliary input during compression.

A central representational choice is conversion from the original single-channel Bayer mosaic to a four-channel RGGB tensor with channels $r$ , $g_1$ , $g_2$ , and $b$ . This preserves the original CFA measurements without demosaicing and exposes inter-channel dependencies within each Bayer cell. The RGGB tensor is then partitioned into non-overlapping patches. During training, raw images are first split into $128\times128$ patches and then randomly cropped to $64\times64$ after augmentation.

The framework is explicitly positioned against two alternatives. First, ordinary image codecs and learned lossless compressors for 8-bit RGB do not naturally match Bayer raw structure or mixed bit-depth regimes. Second, raw reconstruction approaches from sRGB or metadata are inherently lossy and therefore do not preserve the exact raw measurements that are valuable for denoising, super-resolution, and low-light enhancement. RAWIC instead targets direct entropy coding of the original raw data and reconstructs the exact Bayer image.

3. Compression architecture and probabilistic model

RAWIC uses a hyperprior-based ELIC framework as its backbone (Zheng et al., 30 Mar 2026). For each patch $\boldsymbol{x}_i$ , the patch bit depth $\boldsymbol{b}_i$ is embedded to produce $\boldsymbol{e}_i$ , and the analysis transform is conditioned on this embedding:

$r$ 0

The latent representation is quantized, with quantization replaced during training by additive uniform noise. A hyper-analysis transform produces $r$ 1, a hyper-synthesis transform predicts latent distribution parameters, and an overview transform produces prior features $r$ 2. Final symbol probabilities are generated by combining latent-derived prior features with a causal context model.

At the pixel level, RAWIC is a conditional autoregressive entropy model. For a patch $r$ 3, the paper gives the factorization

$r$ 4

and then further factorizes the four-channel RGGB likelihood by channel-autoregressive ordering within each spatial location. This is important because the two green channels and the red and blue values in a Bayer cell are statistically dependent.

Each channel distribution is modeled by a discrete logistic mixture with $r$ 5 components. The entropy parameter network predicts mixture weights, means, scales, and autoregressive coefficients. RAWIC then introduces its central novelty: a bit-depth-adaptive entropy model that masks the predicted PMF to the legal value range implied by the bit depth and renormalizes it. In the paper’s formulation, this prevents the model from allocating probability mass to impossible large values when a patch or symbol has a smaller effective valid range.

The framework also includes channel-autoregressive mean coupling, where later channels depend on previously coded channels through affine adjustments to mixture means. The entropy-coded system is implemented with an arithmetic encoder and arithmetic decoder. The coding objective is the total rate,

$r$ 6

so the total bit cost includes both latent-side information and pixel coding under the adaptive entropy model.

A common misconception is that bit-depth adaptation here is merely camera metadata conditioning. The paper’s formulation is more specific: bit depth is used both as an auxiliary conditioning signal through an embedding and as validity/range information through PMF masking and renormalization. This suggests that RAWIC’s gains are tied to support-aware entropy modeling rather than to camera labels alone.

4. Empirical performance, ablations, and limitations of the compression framework

The reported training data consist of raw images from five cameras in the NUS dataset—Canon 1Ds MkIII, Canon 600D, Olympus EPL6, Panasonic GX1, and Samsung NX2000—plus 5% randomly selected raw images from RAISE (Zheng et al., 30 Mar 2026). The split is 80% training, 10% validation, and 10% testing, with a total training set size of 1088 raw images. Optimization uses Adam, batch size 128, 200 epochs, initial learning rate $r$ 7, a factor-0.1 learning-rate reduction after 10 consecutive epochs without validation improvement, CompressAI, and an NVIDIA A100.

The primary comparison is against traditional lossless codecs: QOI, PNG, WebP, FLIF, JPEG2000, JPEG-LS, and JPEG-XL. The main reported result is an average 7.7% bitrate reduction over JPEG-XL, with the important note that RAWIC’s reported bitrates include the bits for storing bit depths.

Dataset	RAWIC bpp	JPEG-XL bpp
Canon 1Ds MkIII	6.79	7.29
Canon 600D	7.47	7.95
Olympus EPL6	5.11	5.60
Panasonic GX1	5.99	6.57
Samsung NX2000	5.83	6.46
RAISE	7.80	8.29

The ablation on the adaptive entropy model is particularly strong. Replacing the bit-depth-adaptive model with a fixed-bit-depth model degrades performance from 6.79 to 8.78 on Canon 1Ds MkIII, from 7.47 to 8.77 on Canon 600D, from 5.11 to 8.98 on Olympus EPL6, from 5.99 to 9.20 on Panasonic GX1, from 5.83 to 9.37 on Samsung NX2000, and from 7.80 to 9.29 on RAISE. The paper explicitly reports degradations such as +75.7% on Olympus EPL6, +60.7% on Samsung NX2000, and +53.6% on Panasonic GX1. This directly supports the claim that bit-depth adaptation is necessary for raw compression in the proposed setting.

A second notable result is that a single all-in-one model is usually better than separate camera-specific models: 6.79 vs 7.91 on Canon 1Ds MkIII, 7.47 vs 8.85 on Canon 600D, 5.11 vs 6.79 on Olympus EPL6, 5.99 vs 6.46 on Panasonic GX1, 5.83 vs 5.82 on Samsung NX2000, and 7.79 vs 7.80 on RAISE. This contradicts the common expectation that sensor-specific specialization must outperform joint training.

The framework is also retrained on DIV2K for 8-bit RGB lossless compression and is reported at 7.54 bpp on DIV2K, 6.42 bpp on CLIC, and 8.47 bpp on Kodak, slightly outperforming DLPR on those benchmarks. The paper presents this as evidence that the architecture is competitive beyond raw Bayer data, although this is not its main contribution.

The principal limitation is computational cost. For raw-image coding, the reported runtimes are 45.7 / 119.3 sec encode/decode on Canon 600D and 37.1 / 98.4 sec on Olympus EPL6, much slower than JPEG-XL. The method is also tailored to Bayer-pattern input and four-channel RGGB conversion; extension to non-Bayer CFAs such as X-Trans is not demonstrated. The paper explicitly identifies future work on reducing computational complexity and latency.

5. RAWIC as RIS-aided wireless imaging/communications

In RIS-aided sensing, RAWIC refers to reconfigurable intelligent surface aided wireless imaging/communications, represented in the supplied literature by a RIS-aided cooperative ISAC network for imaging-based low-altitude surveillance (Chen et al., 22 Jan 2026). The problem is persistent 3D surveillance of a low-altitude corridor where airborne objects such as UAVs or birds occupy a region of interest above the ground. The central claim is that conventional approaches face deployment-cost, illumination, or signal-strength limits, and that RISs can redirect communication signals into low-altitude space for sensing without requiring the base station to allocate dedicated sensing beams.

The system comprises a TX with $r$ 8 antennas, an RX with $r$ 9 antennas, $g_1$ 0 RISs, and a voxelized 3D ROI. For the $g_1$ 1-th symbol interval, the received signal is modeled as

$g_1$ 2

with RIS phase configurations changing across symbol intervals. The sensing path of interest is

$g_1$ 3

A major distinction is between passive RIS (PRIS) and active RIS (ARIS). PRIS only changes phase and satisfies $g_1$ 4. ARIS permits amplification,

$g_1$ 5

but introduces additional RIS thermal noise and extra power consumption. The paper nevertheless argues that, in the examined imaging regime, ARIS noise that reaches the RX is much weaker than RX noise after propagation, so amplification dominates. Under identical total power constraints, the numerical results show that ARIS outperforms passive RIS, achieving effective imaging and target detection at altitudes up to approximately 300 meters.

The key modeling step is to cast surveillance as imaging. The low-altitude ROI is discretized into voxels with scattering vector $g_1$ 6, sparse because only a few voxels contain targets. For each TX antenna, RIS configuration, RIS, and RX antenna, the cascaded imaging-path coefficient is linear in $g_1$ 7,

$g_1$ 8

which yields the stacked measurement model

$g_1$ 9

The recovery problem is then formulated as sparse reconstruction:

$g_2$ 0

This formulation is intended to avoid error propagation and data association issues associated with conventional delay-angle localization pipelines. Instead of estimating geometric parameters first, the system directly reconstructs a sparse voxel image whose support indicates occupied locations. The paper solves this problem using Subspace Pursuit (SP) and reports average CPU runtimes of 0.075 s for SP versus 0.127 s for OMP. A plausible implication is that the term RAWIC in this literature emphasizes RIS-controlled observation diversity and sparse image inversion more than waveform design.

The RIS-aided imaging formulation is accompanied by a CRLB analysis that makes the geometry-performance dependence explicit (Chen et al., 22 Jan 2026). For unbiased estimation of the $g_2$ 1-th active coefficient under known support $g_2$ 2,

$g_2$ 3

and the average CRLB over the active coefficients is

$g_2$ 4

For $g_2$ 5, the expected per-voxel CRLB is factorized as

$g_2$ 6

with

$g_2$ 7

so the effects of RIS gain $g_2$ 8, inverse noise power $g_2$ 9, and the number of sensing symbols $b$ 0 are separated from geometry terms. The approximate analysis shows that error increases with larger voxel-to-RX distance, larger TX-to-RIS and RIS-to-voxel distances, smaller $b$ 1, smaller $b$ 2, and lower SNR.

These expressions are used to derive deployment rules. The paper concludes that the RX should be placed near the ROI center horizontally, the TX should be close to one RIS, and RIS spacing should balance path loss against aperture size. In the simulation setup, the ROI is $b$ 3 m $b$ 4, discretized into a $b$ 5 grid with $b$ 6, the TX and RX each have 4 antennas, there are 4 RISs with $b$ 7 elements each, the center frequency is 4.9 GHz, $b$ 8, $b$ 9, and the ARIS amplification factor is 40 dB. The reported trends show that ARIS remains effective at 250–300 m altitudes, whereas PRIS degrades much earlier and “hardly achieves effective imaging” above 150 m.

A separate but mathematically relevant compressive-sensing development appears in "Restricted Isometry Random Variables: Probability Distributions, RIC Prediction and Phase Transition Analysis for Gaussian Encoders" (James et al., 2014). That paper generalizes deterministic left and right restricted isometry constants to left and right restricted isometry random variables over an ensemble of random matrices, using the Rayleigh quotient

$128\times128$ 0

as the fundamental object. For i.i.d. Gaussian encoders, the intermediate ratio variable for a fixed support is a central chi-square random variable with $128\times128$ 1 degrees of freedom, the left RIV converges in distribution to a Weibull law, and the right RIV converges to a Gumbel law. The paper argues that eigenvalue-based approaches tend to overestimate the RICs, while the RIV formulation yields more precise estimates and improves RIP-based phase transition analysis. This does not define RAWIC in either of the two contemporary senses above, but it provides a rigorous probabilistic account of sparse-measurement geometry that is directly relevant to compressed-sensing imaging formulations.

Taken together, these lines of work show that RAWIC currently names two different research programs. One is a specialized neural lossless codec for Bayer raw data whose main technical idea is bit-depth-adaptive entropy modeling. The other is a RIS-assisted sparse imaging paradigm in which programmable propagation, compressed sensing, and geometry-aware CRLB analysis are used to make low-altitude wireless imaging viable. Their commonality lies in explicit modeling of structured measurement support, adaptive probability assignment, and the use of mathematically constrained inference rather than generic fixed-format processing.