Measurement-Aware Cross Attention (MACA)

• The paper demonstrates that MACA improves reconstruction quality (up to +0.35 dB PSNR) by integrating measurement-guided queries and keys into both forward and backward attention stages.
• MACA employs a two-stage process where Forward Attention aggregates down-sampled features and Backward Attention redistributes them to full resolution, ensuring measurement-aware feature fusion.
• Integrated within MEUNet alongside AKCS and a U-Net denoiser, MACA delivers faster, more accurate image reconstruction with reduced computational overhead compared to full-resolution self-attention.

Measurement-Aware Cross Attention (MACA) is a cross-attention mechanism introduced in “Physics-guided Deep Unfolding Network for Enhanced Kronecker Compressive Sensing” to learn implicit measurement representations during image compressed sensing reconstruction (Qu et al., 13 Aug 2025). It is designed for the setting in which a high-fidelity image is reconstructed from its compressed measurement, and it operates by coupling feature space and measurement space through two sub-stages, “Forward Attention” and “Backward Attention.” Within the paper’s measurement-enhanced unfolding network (MEUNet), MACA is integrated with asymmetric Kronecker CS (AKCS) and a U-Net style deep denoiser so that the compressed measurement explicitly guides feature aggregation and redistribution across reconstruction stages.

1. Formal definition

MACA takes as inputs the current stage’s feature map and the 2D compressive measurement. The paper defines $F_{\rm in}\in\mathbb R^{N_H\times N_W\times N_C}$ as the current stage’s feature map, $Y\in\mathbb R^{N_h\times N_w}$ as the 2D compressive measurement, and $D$ as a down-sampling factor, with average pooling by $D\times D$ (Qu et al., 13 Aug 2025). The mechanism proceeds in two sub-stages.

In the measurement encoding step, a small CNN called the Measurement Encoder Block (MEB) produces both a Query bank and a Key bank from the measurement:

$$Q_Y,\;K_Y=\operatorname{MEB}(Y)\in\mathbb R^{N_Y\times d_k}.$$

The paper states that each bank has length $N_Y$ and embedding dimension $d_k$, often with $N_Y=N_h\,N_w$.

The first sub-stage, “Forward Attention,” aggregates information from the feature map into a compact representation. The feature map is down-sampled and linearly embedded:

$$\tilde F = \operatorname{AvgPool}_{D}(F_{\rm in}) \in\mathbb R^{\tfrac{N_H}{D}\times\tfrac{N_W}{D}\times d_k}.$$

After flattening into $K_D,\,V_D\in\mathbb R^{N_D\times d_k}$ with $Y\in\mathbb R^{N_h\times N_w}$0, the low-resolution attention is computed as

$Y\in\mathbb R^{N_h\times N_w}$1

The paper characterizes this step by stating that each measurement query in $Y\in\mathbb R^{N_h\times N_w}$2 “asks” which down-sampled feature patch is most relevant, and $Y\in\mathbb R^{N_h\times N_w}$3 collects these pooled answers.

The second sub-stage, “Backward Attention,” propagates the compact representation back to full resolution. The full-resolution feature map is linearly embedded into queries

$Y\in\mathbb R^{N_h\times N_w}$4

and then high-resolution attention is computed as

$Y\in\mathbb R^{N_h\times N_w}$5

The aggregated answers are then gathered back to every pixel:

$Y\in\mathbb R^{N_h\times N_w}$6

after which $Y\in\mathbb R^{N_h\times N_w}$7 is reshaped into $Y\in\mathbb R^{N_h\times N_w}$8 and fused back into the network’s feature pathway.

2. Motivation and “measurement awareness”

The paper distinguishes MACA from standard cross-attention by stating that standard cross-attention, such as in ViT, uses queries and keys drawn from the same feature stream or from image–image and image–text pairings, whereas MACA makes one side of the attention—“the queries and keys in both forward and backward passes”—come exclusively from the compressed measurement (Qu et al., 13 Aug 2025). In this sense, the module is “measurement-aware” because the measurement domain directly controls both the low-resolution aggregation and the high-resolution redistribution.

The stated intuition is that the network learns how each spatial feature patch should be weighted in the forward pass and then how that global summary should be redistributed in the backward pass, all under direct guidance of the measurement domain. The paper further states that this “measurement awareness” ensures that the global prior contained in the low-dimensional measurement, which encodes information about the entire image via the learned sensing matrix, is explicitly and adaptively fused into every pixel’s representation during reconstruction.

A second motivation concerns informative representations. The paper argues that by first compressing the image features and then attentively re-projecting via the measurement queries, MACA enforces the network to focus on those components of the feature map that actually matter to the measurement. In contrast, a vanilla self-attention among image tokens might attend to patterns that the measurement cannot observe. No closed-form proof is given in the paper, but the reported ablation indicates that removing MACA costs approximately $Y\in\mathbb R^{N_h\times N_w}$9 dB PSNR and approximately $D$0 SSIM. This suggests that the mechanism contributes measurement-driven information beyond local feature refinement alone.

3. Placement within MEUNet and unfolded reconstruction

MACA is not presented as an isolated module; it is part of MEUNet, which unfolds ISTA under the AKCS model for $D$1 stages (Qu et al., 13 Aug 2025). Each stage alternates between a gradient descent step and a deep denoiser.

The gradient descent step is defined as

$D$2

The deep denoiser is then applied:

$D$3

where $D$4 is a U-Net style network with Convolution-Transformer blocks (CTB) and MACA blocks in its encoder at each scale. The paper states that all $D$5 parameters are shared across stages.

Within the denoiser, MACA is inserted after the local Convolution-Transformer block at each encoder level. The stated workflow is: in each encoder level of $D$6, after the local CTB, one MACA is inserted; then the network down-samples and repeats. Decoder levels invert the process via up-sampling and skip connections, and the paper explicitly notes that there is no MACA in the decoder.

The paper provides the following stage-level procedure:

$d_k$7

This organization places MACA precisely at the interface between local feature processing and multiscale encoding, rather than as a generic attention layer applied uniformly throughout the network.

4. Relation to asymmetric Kronecker compressive sensing

The paper introduces MACA together with asymmetric Kronecker CS, and their interaction is explicit rather than incidental (Qu et al., 13 Aug 2025). Standard Kronecker CS is written as

$D$7

AKCS breaks this symmetry by allowing each row-measurement to use its own column-basis:

$D$8

with vectorized form

$D$9

The paper reports coherence bounds for both models. For standard KCS,

$D\times D$0

while for AKCS, Theorem 1 gives

$D\times D$1

The paper states that this is strictly smaller than $D\times D$2 whenever $D\times D$3.

The stated interaction with MACA is that lower-coherence measurements cause the compressed vector $D\times D$4 to carry “more uniformly distributed information” about the image. MACA can therefore extract a richer set of queries $D\times D$5 that are less biased toward a few correlated directions, improving the global guidance it provides to the feature stream. This is a direct claim in the exposition rather than a separate theoretical theorem for MACA itself. A plausible implication is that MACA’s efficacy depends not only on the attention design but also on the statistical properties of the sensing model that generates the measurement tokens.

5. Empirical behavior and reported gains

The paper attributes part of MEUNet’s reconstruction improvement to MACA and reports both end-to-end and ablation results (Qu et al., 13 Aug 2025). On Set11 at SR=10% (Table 1), ProxUnroll, identified as the best prior in the comparison, achieves $D\times D$6 dB/$D\times D$7, while MEUNet with AKCS+MACA achieves $D\times D$8 dB/$D\times D$9, corresponding to $$Q_Y,\;K_Y=\operatorname{MEB}(Y)\in\mathbb R^{N_Y\times d_k}.$$0 dB and $$Q_Y,\;K_Y=\operatorname{MEB}(Y)\in\mathbb R^{N_Y\times d_k}.$$1 SSIM.

The runtime-versus-accuracy comparison in Table 4 reports the following values.

Method	Params / GFlops / Time	PSNR
ProxUnroll	3.90 M / 107.7 GFlops / 0.27 s	32.55 dB
MEUNet	2.59 M / 141.3 GFlops / 0.21 s	32.90 dB

The paper states that MEUNet is both faster ($$Q_Y,\;K_Y=\operatorname{MEB}(Y)\in\mathbb R^{N_Y\times d_k}.$$2) and more accurate ($$Q_Y,\;K_Y=\operatorname{MEB}(Y)\in\mathbb R^{N_Y\times d_k}.$$3 dB) despite fewer parameters.

The MACA-specific ablation in § 4.3 removes MACA while keeping AKCS and CTB, yielding

$$Q_Y,\;K_Y=\operatorname{MEB}(Y)\in\mathbb R^{N_Y\times d_k}.$$4

The paper interprets this as confirmation of MACA’s unique contribution beyond the improved sensing model and local Transformer features. Because no closed-form proof is given, the empirical evidence is the primary basis for attributing this effect to implicit measurement representations.

6. Complexity, hyperparameters, and implementation details

The paper presents MACA as a lightweight module relative to naïve full-resolution self-attention, while still imposing a measurable overhead (Qu et al., 13 Aug 2025). The forward attention cost is reported as

$$Q_Y,\;K_Y=\operatorname{MEB}(Y)\in\mathbb R^{N_Y\times d_k}.$$5

and the backward attention cost as

$$Q_Y,\;K_Y=\operatorname{MEB}(Y)\in\mathbb R^{N_Y\times d_k}.$$6

Because $$Q_Y,\;K_Y=\operatorname{MEB}(Y)\in\mathbb R^{N_Y\times d_k}.$$7 typically is $$Q_Y,\;K_Y=\operatorname{MEB}(Y)\in\mathbb R^{N_Y\times d_k}.$$8–$$Q_Y,\;K_Y=\operatorname{MEB}(Y)\in\mathbb R^{N_Y\times d_k}.$$9, and $N_Y$0, the combined cost is stated to be far below a naïve full-resolution self-attention of order $N_Y$1. In practice, the paper reports that MACA adds approximately $N_Y$2–$N_Y$3 overhead in GFlops to the CTB-only denoiser but unlocks the $N_Y$4 dB gain alone, and $N_Y$5 dB when combined with AKCS.

The hyperparameters reported for MACA and the unfolded model are specific. The embedding dimension in MACA is $N_Y$6, set equal to the channel dimension after CTB. The number of heads in the measurement-feature attention is not explicitly tuned; practically a single-head with dimension $N_Y$7 is used to save memory. The down-sampling factor is $N_Y$8 in all experiments, and the number of unfolding stages is $N_Y$9.

The training protocol is also specified. The training set consists of $d_k$0 images from BSD500, augmented to $d_k$1 patches. Optimization uses Adam with $d_k$2. The schedule is $d_k$3 epochs at $d_k$4 followed by $d_k$5 epochs at $d_k$6, and the loss is mean squared error between reconstruction and ground truth.

7. Conceptual significance and common points of interpretation

MACA is presented as a mechanism for learning implicit measurement representations, and the paper contrasts this with the claim that unfolding networks derive superiority over non-unfolding ones from sufficient gradient descents, which are termed explicit measurement representations (Qu et al., 13 Aug 2025). In that framing, MACA extends the role of the measurement beyond the gradient descent step by reintroducing measurement-conditioned information inside the learned denoiser. This suggests a division of labor in which explicit measurement representations arise from the physics-guided update, while implicit measurement representations arise from attention-mediated feature modulation.

A common point of interpretation is whether MACA is simply another Transformer block. The paper’s description argues against that reading: MACA is defined by the fact that the compressed measurement exclusively provides the measurement-side queries and keys in both forward and backward passes. Its purpose is therefore not generic long-range dependency modeling, but measurement-guided aggregation and redistribution.

Another point is the status of its empirical support. The paper explicitly notes that no closed-form proof is given for the claim that MACA learns more informative representations, and the evidence offered is the ablation result showing degradation when MACA is removed. An objective reading is therefore that MACA is empirically validated within the MEUNet framework, while its representational role is motivated by architectural design and comparative performance rather than by a separate standalone theorem.

In summary, MACA is a two-step aggregate-in-low, propagate-in-high attention module that tightly couples measurement space and feature space. Within MEUNet, it functions as the mechanism by which compressed measurements provide global guidance to multiscale feature reconstruction, complementing AKCS at the sensing phase and unfolded gradient descent at the optimization phase.