MEUNet: Enhanced Unfolding for CS

• The paper introduces MEUNet, a deep unfolding network that integrates Asymmetric Kronecker Compressed Sensing (AKCS) and Measurement-Aware Cross Attention (MACA) to enhance image reconstruction.
• MEUNet employs explicit gradient-descent updates and implicit attention-based fusion to leverage measurement information, leading to superior PSNR and SSIM performance.
• MEUNet demonstrates higher reconstruction accuracy and faster inference with a smaller model size by co-optimizing sensing design and reconstruction methodology.

Searching arXiv for the focal paper and closely related compressed sensing unfolding methods to attach accurate arXiv citations. Measurement-Enhanced Unfolding Network (MEUNet) is a deep unfolding network for image compressed sensing reconstruction that is designed to address two measurement bottlenecks: suboptimal coherent measurements at the sensing phase and underutilization of measurement information at the reconstruction phase. It is introduced in “Physics-guided Deep Unfolding Network for Enhanced Kronecker Compressive sensing” (Qu et al., 13 Aug 2025), where it is instantiated by integrating Asymmetric Kronecker Compressed Sensing (AKCS) and Measurement-Aware Cross Attention (MACA) into a widely used unfolding architecture. In the formulation presented there, MEUNet targets both explicit measurement representations, associated with sufficient gradient descents in unfolding, and implicit measurement representations, learned through attention-based fusion of measurement information, with the stated goal of improving reconstruction accuracy and inference speed (Qu et al., 13 Aug 2025).

1. Concept and problem setting

MEUNet is situated in image compressed sensing (CS), where the task is to reconstruct a high-fidelity image from compressed measurements. The motivating diagnosis is twofold. First, existing methods are described as deficient in incoherent compressed measurement at sensing phase because structured sensing matrices induce suboptimal coherence properties. Second, prior approaches are described as relying on implicit measurement representations at reconstruction phase, thereby limiting the use of measurement information during deep reconstruction (Qu et al., 13 Aug 2025).

The architecture addresses these two issues through two named components. AKCS modifies the sensing model in order to reduce mutual coherence relative to classic Kronecker compressed sensing. MACA introduces an explicit mechanism for injecting measurement-conditioned information into the reconstruction process. Within the paper’s framing, these two components are complementary: AKCS improves the quality and diversity of the measurements themselves, while MACA improves how those measurements are exploited once reconstruction begins (Qu et al., 13 Aug 2025).

A plausible implication is that MEUNet should be understood not merely as a denoising-centric unrolled network, but as a measurement-centric variant of deep unfolding in which sensing design and iterative reconstruction are co-optimized conceptually, even if the paper presents them as architectural modules rather than as a single end-to-end analytical theory.

2. Architectural composition

The backbone of MEUNet is a Deep Unfolding Network (DUN) with $K$ iterative stages, each stage mimicking a gradient descent update for data fidelity followed by a trainable deep denoiser (Qu et al., 13 Aug 2025). This places MEUNet within the standard unrolled optimization paradigm, but its distinguishing feature is that measurement structure is explicitly incorporated at both the sensing and denoising levels.

The overall pipeline is described as follows. First, forward sensing is performed with AKCS, producing compressed measurements. Second, unfolding reconstruction proceeds stage by stage. At each stage, the network applies a gradient-based data fidelity projection and then denoises or interprets the intermediate reconstruction using a deep denoiser that leverages both the measurement and the previous stage output, including MACA (Qu et al., 13 Aug 2025).

The denoiser is a symmetric UNet with multi-scale feature learning. Each layer includes Convolution-Transformer blocks and MACA blocks. In this design, the Convolution-Transformer blocks provide local and global modeling, while MACA acts as a plug-in mechanism for measurement-aware feature fusion. The resulting denoiser therefore combines multi-scale encoder–decoder processing, token-style global interactions, and explicit measurement-conditioned attention in each layer of the reconstruction network (Qu et al., 13 Aug 2025).

This suggests that MEUNet departs from earlier unfolding models that treat the measurement primarily as an input to the data fidelity term. Here, measurement information is also routed into the learned proximal component, so the denoiser is not measurement-agnostic.

3. Asymmetric Kronecker Compressed Sensing

AKCS is introduced to overcome a stated limitation of classic Kronecker CS. In the classic model, the sensing matrix has the form $\Phi = \Phi_1 \otimes \Phi_2$, and this highly structured form is said to increase mutual coherence because columns become too similar, limiting the expressivity of the measurement operator and the quality of reconstruction (Qu et al., 13 Aug 2025).

For a 2D image $\mathbf{X} \in \mathbb{R}^{H \times W}$ and measurement $\mathbf{Y} \in \mathbb{R}^{m \times n}$, classic KCS is written as

$\mathbf{Y} = \Phi_1 \mathbf{X} \Phi_2^\top.$

AKCS replaces the global Kronecker product with row-adaptive sub-matrices. For each row $i$,

$$\text{row}_i(\mathbf{Y}) = \mathbf{a}_i \mathbf{X} \mathbf{B}_i^\top,$$

where $\mathbf{a}_i \in \mathbb{R}^{1 \times H}$ and $\mathbf{B}_i \in \mathbb{R}^{n \times W}$, and the full measurement is

$$\mathbf{Y} = \mathcal{C}_{i=1}^m [\mathbf{a}_i \mathbf{X} \mathbf{B}_i^\top].$$

The key idea is that each measurement row is uniquely parameterized, breaking column and row symmetry (Qu et al., 13 Aug 2025).

The paper presents a coherence analysis using the definition

$\Phi = \Phi_1 \otimes \Phi_2$0

where $\Phi = \Phi_1 \otimes \Phi_2$1 is the $\Phi = \Phi_1 \otimes \Phi_2$2th normalized column. For classic KCS,

$\Phi = \Phi_1 \otimes \Phi_2$3

and with i.i.d. Gaussian random matrices,

$\Phi = \Phi_1 \otimes \Phi_2$4

For AKCS, with i.i.d. Gaussian submatrices, Theorem 1 gives

$\Phi = \Phi_1 \otimes \Phi_2$5

with $\Phi = \Phi_1 \otimes \Phi_2$6. The paper emphasizes that because $\Phi = \Phi_1 \otimes \Phi_2$7, AKCS substantially lowers coherence compared to KCS (Qu et al., 13 Aug 2025).

The stated theoretical impact is that reduced coherence increases measurement diversity and lowers the ill-posedness of the inverse problem, enabling higher-quality reconstructions. The paper also states that this coherence improvement yields better restricted isometry properties and stronger reconstruction guarantees. A plausible implication is that AKCS functions as a sensing-side regularization mechanism: rather than relying exclusively on a stronger reconstructor, it changes the inverse problem itself.

4. Unfolding dynamics and explicit measurement representations

MEUNet adopts an ISTA-style update adapted to AKCS. At stage $\Phi = \Phi_1 \otimes \Phi_2$8, the data fidelity projection is

$\Phi = \Phi_1 \otimes \Phi_2$9

where $\mathbf{X} \in \mathbb{R}^{H \times W}$0 is a learnable step size and $\mathbf{X} \in \mathbb{R}^{H \times W}$1 is measurement row $\mathbf{X} \in \mathbb{R}^{H \times W}$2. The proximal mapping is implemented by a deep denoiser,

$\mathbf{X} \in \mathbb{R}^{H \times W}$3

where $\mathbf{X} \in \mathbb{R}^{H \times W}$4 denotes the deep denoising network and may be shared across stages (Qu et al., 13 Aug 2025).

The paper explicitly argues that unfolding networks outperform non-unfolding ones because of sufficient gradient descents, referring to these as explicit measurement representations. In this interpretation, the iterative data-fidelity step is not merely a numerical update but a mechanism for preserving measurement-aligned structure through the sequence of stages (Qu et al., 13 Aug 2025).

This view is important for interpreting MEUNet’s design. The contribution is not limited to replacing a proximal operator with a stronger denoiser. Rather, the method distinguishes two representational channels. One is explicit and arises from the gradient descent–like update that directly enforces consistency with measurements. The other is implicit and is learned through MACA. The architecture is therefore framed as an overview of optimization-based and attention-based measurement usage.

5. Measurement-Aware Cross Attention

MACA is introduced to incorporate more implicit measurement information into reconstruction beyond what is already used in the data fidelity term (Qu et al., 13 Aug 2025). The stated motivation is to learn measurement priors that standard unfolding updates do not capture.

The mechanism has two stages. In the forward, low-resolution attention stage, feature maps are downsampled to limit computation. Measurement-encoded queries and keys, denoted $\mathbf{X} \in \mathbb{R}^{H \times W}$5 and $\mathbf{X} \in \mathbb{R}^{H \times W}$6, are used to aggregate over downsampled image features $\mathbf{X} \in \mathbb{R}^{H \times W}$7: $\mathbf{X} \in \mathbb{R}^{H \times W}$8 This produces a set of global answer vectors of size $\mathbf{X} \in \mathbb{R}^{H \times W}$9 (Qu et al., 13 Aug 2025).

In the backward, high-resolution attention stage, the model projects this information back to the spatial domain. Each high-resolution pixel location has its own query vector and computes an attention map with measurement-derived keys: $\mathbf{Y} \in \mathbb{R}^{m \times n}$0 These attention weights are then used to propagate the global answers across all spatial positions. The paper summarizes this as an “Aggregate-in-Low, Propagate-in-High” strategy, intended to preserve tractability while injecting measurement-conditioned global context into the reconstruction stream at every stage of the UNet (Qu et al., 13 Aug 2025).

MACA is inserted as a plug-in block within each encoder and decoder layer of the deep denoiser. Its role is therefore pervasive rather than peripheral. It does not operate only once at the network entrance or exit; it modulates multi-scale feature processing throughout the denoiser hierarchy (Qu et al., 13 Aug 2025).

The paper describes MACA as the first explicit measurement-aware attention in a DUN. Within the confines of the reported framework, the significance of this claim is that measurement information becomes a first-class conditioning signal for feature aggregation, rather than being confined to the observation model alone. A plausible implication is that MACA is particularly relevant at low sampling ratios, where weak or ambiguous spatial evidence may benefit from global measurement-conditioned cues.

6. Empirical results and ablations

The reported experiments include synthetic data on Set11 and BSD500 and real data from single-pixel imaging experiments (Qu et al., 13 Aug 2025). On synthetic benchmarks, MEUNet is compared with prior DUNs and related models including ReconNet, ISTA-Net+, CSNet+, SCSNet, AMP-Net, TransCS, SAUNet, HATNet, ProxUnroll, and CPPNet.

For the sampling ratio $\mathbf{Y} \in \mathbb{R}^{m \times n}$1, the paper reports the following PSNR and SSIM values:

Method	PSNR	SSIM
ReconNet	24.38	0.7301
ISTA-Net+	26.64	0.8087
CSNet+	28.34	0.8580
AMP-Net	29.43	0.8880
TransCS	29.54	0.8877
DGUNet+	30.93	0.9088
SAUNet	32.15	0.9147
HATNet	32.26	0.9182
ProxUnroll	32.55	0.9226
MEUNet	32.90	0.9270

The qualitative comparison is summarized in the paper as showing sharper edges and better texture preservation, especially at low sampling ratios such as $\mathbf{Y} \in \mathbb{R}^{m \times n}$2–$\mathbf{Y} \in \mathbb{R}^{m \times n}$3 (Qu et al., 13 Aug 2025). On real data, MEUNet is reported to outperform recent KCS-based state-of-the-art methods, specifically HATNet and ProxUnroll, on physically acquired data, which the paper interprets as evidence of robustness and real-world feasibility (Qu et al., 13 Aug 2025).

The ablation study identifies AKCS as the most consequential component among those removed individually. Removing AKCS leads to the largest drop in performance, reported as $\mathbf{Y} \in \mathbb{R}^{m \times n}$4 dB PSNR, followed by removing Transformer, Convolution, or MACA blocks (Qu et al., 13 Aug 2025). Within the article’s internal logic, this emphasizes that improved sensing is not an auxiliary detail but a primary driver of the overall gain.

7. Efficiency, significance, and scope

The paper reports model size, computational cost, and inference speed for MEUNet and several comparators. MEUNet is listed as having $\mathbf{Y} \in \mathbb{R}^{m \times n}$5M parameters, $\mathbf{Y} \in \mathbb{R}^{m \times n}$6G FLOPs, and $\mathbf{Y} \in \mathbb{R}^{m \times n}$7s/image inference. The corresponding values given for ProxUnroll are $\mathbf{Y} \in \mathbb{R}^{m \times n}$8M, $\mathbf{Y} \in \mathbb{R}^{m \times n}$9G, and $\mathbf{Y} = \Phi_1 \mathbf{X} \Phi_2^\top.$0s, while HATNet is reported as $\mathbf{Y} = \Phi_1 \mathbf{X} \Phi_2^\top.$1M, $\mathbf{Y} = \Phi_1 \mathbf{X} \Phi_2^\top.$2G, and $\mathbf{Y} = \Phi_1 \mathbf{X} \Phi_2^\top.$3s (Qu et al., 13 Aug 2025). The paper’s conclusion is that MEUNet is smaller and faster while also being more accurate on the reported tasks.

The significance of MEUNet, as framed by the source, lies in the coupling of two ideas. AKCS addresses the measurement operator itself by lowering mutual coherence and improving sensing diversity with negligible additional computational cost. MACA addresses the representational pathway by learning implicit measurement priors through explicit measurement-aware attention. Their combination is presented as yielding state-of-the-art results for both simulated and real data in compressed sensing reconstruction (Qu et al., 13 Aug 2025).

Several interpretive points follow directly from this presentation. First, MEUNet treats the sensing phase and reconstruction phase as jointly consequential; reconstruction performance is not attributed solely to a stronger decoder. Second, the method positions measurement information as both a constraint and a feature source. Third, the reported gains in parameter count and runtime indicate that the design is not framed as accuracy-at-any-cost, but as an efficiency-conscious architecture within deep unfolding.

A potential misconception is that MEUNet is only a denoiser improvement over earlier unrolled methods. The source material does not support that characterization. Its central claims are instead built around a new asymmetric sensing model, a theory-driven coherence argument, and a measurement-aware attention mechanism integrated into the denoiser. Another potential misconception is that its advantage arises only from transformerization. The ablation summary counters this by identifying AKCS as causing the largest performance drop when removed (Qu et al., 13 Aug 2025).

In the paper’s own synthesis, MEUNet advances compressed sensing reconstruction by breaking sensing matrix limitations through AKCS, innovatively fusing measurement signals through MACA, and integrating both into a DUN that is reported to achieve state-of-the-art reconstruction accuracy together with favorable parameter count and inference speed (Qu et al., 13 Aug 2025).