---
title: Deep Equilibrium Convolutional Sparse Coding
url: https://www.emergentmind.com/topics/deep-equilibrium-convolutional-sparse-coding-decsc
type: topic
---

# Deep Equilibrium Convolutional Sparse Coding

Searching arXiv for DECSC and closely related deep equilibrium sparse coding papers.
Deep Equilibrium Convolutional Sparse Coding (DECSC) denotes an implicit, optimization-grounded class of models in which convolutional sparse coding (CSC) inference is formulated as a fixed-point problem and solved within a deep equilibrium (DEQ) framework. In its explicit form, DECSC is instantiated for hyperspectral image (HSI) denoising by a model that unifies shared 2D convolutional sparse representation, unshared 3D convolutional sparse representation, nonlocal spatial self-similarity, and detail-preserving regularization inside an infinite-depth equilibrium architecture [2508.15553]. The resulting formulation retains the physical interpretability of sparse coding while replacing finite unrolling depth by equilibrium computation.

## 1. Problem setting and defining characteristics

DECSC is developed for HSI denoising under the observation model
$$
Y = X + N,
$$
where $Y \in \mathbb{R}^{H \times W \times B}$ is the noisy HSI, $X$ is the unknown clean HSI, and $N$ is additive noise. The method is designed for noise regimes that include non-i.i.d. Gaussian noise, mixture noise, and spectrally correlated variance noise, with the aim of preserving the physical properties of HSIs: consistent spatial structures across bands, meaningful spectral reflectance behavior, and fine local spatial-spectral detail [2508.15553].

A defining feature of DECSC is the decomposition
$$
X = C + U,
$$
where $C$ denotes **global inter-band common (GIC)** structures and $U$ denotes **local spatial-spectral unique (LSU)** structures. This decomposition is not an auxiliary narrative device but the central modeling principle. The GIC component encodes the fact that scene objects maintain largely consistent spatial layouts across wavelengths, whereas the LSU component captures finer band-dependent and local spatial-spectral variations [2508.15553].

The term “DECSC” is most precise when three properties are present simultaneously: an explicit CSC model with convolutional dictionaries and sparse coefficients, a DEQ fixed-point formulation of the iterative inference map, and learned regularization embedded inside the equilibrium operator. In this sense, DECSC is more specific than generic DEQ-based inverse-problem reconstruction and more specific than finite-depth learned CSC.

## 2. Intellectual lineage and delimitation

DECSC sits at the intersection of CSC, learned sparse inference, and DEQ. Its technical lineage can be organized as follows.

| Work | Main contribution | Relation to DECSC |
|---|---|---|
| "Optimization Methods for Convolutional Sparse Coding" [1406.2407] | Canonical CSC objective and solver analysis | Optimization backbone, no DEQ |
| "Learned Convolutional Sparse Coding" [1711.00328] | Convolutional LISTA-style recurrent encoder and convolutional decoder | Unrolled learned CSC precursor |
| "Connections between Deep Equilibrium and Sparse Representation Models with Application to Hyperspectral Image Denoising" [2203.15901] | DEQ reformulation of sparse coding with a matrix dictionary and CNN prior | Deep equilibrium sparse coding, not convolutional CSC |
| "Deep Equilibrium Models for Video Snapshot Compressive Imaging" [2201.06931] | DEQ machinery for inverse problems with fixed-point solving and implicit differentiation | Methodological DEQ template, not CSC |
| "MsDC-DEQ-Net: Deep Equilibrium Model (DEQ) with Multi-scale Dilated Convolution for Image Compressive Sensing (CS)" [2401.02884] | DEQ-ISTA-inspired CS reconstruction with convolutional proximal modules | Not a canonical DECSC method |
| "Deep Equilibrium Convolutional Sparse Coding for Hyperspectral Image Denoising" [2508.15553] | Explicit DECSC framework for HSI denoising | Canonical DECSC instance |

The canonical CSC formulation predates DEQ. In the standard multi-filter setting, CSC models the signal as
$$
x \approx \sum_j d_j * z_j,
$$
and solves an $\ell_1$-regularized convolutional synthesis problem with explicit sparse coefficient maps and convolutional atoms [1406.2407]. This is the optimization backbone that DECSC inherits.

Learned CSC introduced a tied-weight convolutional recurrent inference map of LISTA/ISTA type,
$$
\mathbf{z}_{k+1}= S_\theta\!\left( \mathbf{z}_k + \mathbf{w_e} * (\mathbf{x} - \mathbf{w_d}*\mathbf{z}_k) \right),
$$
together with a linear convolutional decoder. That work already supplied a natural fixed-point candidate map for CSC, but it remained a finite unrolled model with $K=3$ recurrent steps rather than an equilibrium solve [1711.00328].

Deep equilibrium sparse coding then emerged in non-convolutional form. In hyperspectral denoising, a blockwise sparse coding problem with matrix dictionary $D$ and a CNN prior was rewritten as a DEQ on the sparse coding matrix $G$, creating an explicit bridge between sparse representation theory and equilibrium models, but not a convolutional sparse coding framework [2203.15901].

By contrast, DEQ inverse-problem papers such as the video SCI formulation and the ISTA-inspired compressive sensing reconstruction model showed how to reinterpret iterative reconstruction maps as equilibrium operators, solve them with Anderson acceleration, and train them with implicit differentiation or related machinery, but without explicit convolutional sparse codes or convolutional dictionaries [2201.06931], [2401.02884].

A common misconception is therefore that any DEQ model with convolutions and soft-thresholding qualifies as DECSC. Strictly, DECSC requires explicit CSC structure. The distinction is important because DECSC defines equilibrium over sparse coefficient variables tied to convolutional synthesis operators, rather than merely over reconstructed images.

## 3. Core mathematical formulation

The DECSC model for HSI denoising contains two CSC branches.

The **shared 2D CSC branch** represents the GIC component band by band:
$$
C_b = K_b \star S = \sum_{m=1}^{M} k_{b,m} \star S_m,
$$
where $K_b = \{k_{b,m}\}_{m=1}^{M}$ is the 2D convolutional dictionary for band $b$, and $S = \{S_m\}_{m=1}^{M}$ is the sparse code shared across all bands. The crucial modeling choice is that the coefficients are shared over the spectral dimension, while the filters remain band-specific [2508.15553].

The **unshared 3D CSC branch** represents the LSU component by 3D convolution:
$$
U = D \star H = \sum_{j=1}^{J} d_j \star h_j,
$$
where $D = \{d_j\}_{j=1}^{J}$ is a 3D convolutional dictionary and $H = \{h_j\}_{j=1}^{J}$ are the corresponding sparse coefficients [2508.15553].

Without learned regularizers, the joint CSC objective is
$$
\min_{S, H}
\frac{1}{2} \left\| Y - K \otimes S - D \star H \right\|_F^2
+ \lambda_1 \|S\|_1 + \lambda_2 \|H\|_1.
$$
The full formulation augments this with learned regularization terms:
$$
\begin{split}
\min_{S, H} \quad &
\frac{1}{2} \left\| Y - K \otimes S - D \star H \right\|_F^2
+ \lambda_1 \|S\|_1 + \lambda_2 \|H\|_1 \\
&+ \mu_1 \mathcal{R}(S) + \mu_2 \mathcal{R}(H).
\end{split}
$$
Here $\mathcal{R}(S)$ encodes nonlocal spatial self-similarity in the shared GIC code, and $\mathcal{R}(H)$ encodes detail-preserving regularization in the LSU code [2508.15553].

The alternating subproblems are
$$
\min_{S} \;\frac{1}{2} \left\| Y - K \otimes S - D \star H \right\|_F^2 + \lambda_1 \|S\|_1 + \mu_1 \mathcal{R}(S),
$$
and
$$
\min_{H} \;\frac{1}{2} \left\| Y - K \otimes S - D \star H \right\|_F^2 + \lambda_2 \|H\|_1 + \mu_2 \mathcal{R}(H).
$$
After estimating $S$ and $H$, the clean HSI is reconstructed as
$$
\widehat{X} = K \otimes S + D \star H.
$$

This formulation places DECSC squarely inside the canonical CSC tradition: it contains explicit sparse coefficients, explicit convolutional dictionaries, explicit synthesis operators, and $\ell_1$ sparsity penalties. What distinguishes it from earlier CSC is the equilibrium treatment of the iterative solver and the insertion of learned regularizers inside that solver.

## 4. Fixed-point construction and equilibrium training

DECSC converts proximal-gradient sparse coding updates into a DEQ layer. The branchwise updates are
$$
\begin{split}
S^{(t+1)} &= \text{Net}_1\!\left(
\text{Soft}_{\theta_1}\!\left(
S^{(t)} + K^T \otimes \left(Y - K\otimes S^{(t)} - D \star H^{(t)} \right)
\right)\right), \\
H^{(t+1)} &= \text{Net}_2\!\left(
\text{Soft}_{\theta_2}\!\left(
H^{(t)} + D^T \star \left(Y - K\otimes S^{(t+1)} - D\star H^{(t)}\right)
\right)\right).
\end{split}
$$
The soft-thresholding operator is
$$
\text{Soft}_{\theta}(x)=\text{sign}(x)\cdot\max(|x|-\theta, 0).
$$
Each update therefore has the classical sparse-coding structure: a gradient backprojection through transposed convolutions, an $\ell_1$ proximal map, and a learned regularization operator [2508.15553].

The paper abstracts this to a common DEQ layer:
$$
\alpha^{(t+1)} =
\text{Net}\left(\text{Soft}_{\theta} \left( \alpha^{(t)} + E^{T} \star \left( z - E \star \alpha^{(t)} \right) \right)\right),
$$
with learnable map
$$
f_{\Theta}(\alpha^{(t)}, z)=
\text{Net}\left(\text{Soft}_{\theta}\left( \alpha^{(t)} + W_E \star \left( z - E \star \alpha^{(t)} \right) \right)\right),
$$
where $\Theta=\{\theta, W_E, E\}$ and $W_E = E^T$. The equilibrium condition is
$$
\alpha^{*}= f_{\Theta}\left ( \alpha^*, z \right ).
$$

This fixed-point view turns the model into an implicit infinite-depth network. In a finite unrolled architecture, one chooses a depth $T$ and outputs $\alpha^{(T)}$. In DECSC, the output is instead the equilibrium state $\alpha^*$, assuming convergence. This aligns the network definition with the optimization objective more directly than fixed-depth unfolding.

The forward pass uses **Anderson acceleration**:
$$
\alpha^{(t+1)}=
(1-\beta)\sum_{i=0}^{m} \gamma_i^{(t)} \alpha^{(t-i)}
+\beta \sum_{i=0}^{m}\gamma_i^{(t)} f_{\Theta}(\alpha^{(t-i)}, z),
$$
with residual
$$
g_{\Theta}(\alpha^{(t)}, z)=f_{\Theta}(\alpha^{(t)}, z)-\alpha^{(t)},
$$
and coefficients obtained from
$$
\arg \min_{\gamma}\|G\gamma\|_2^2, \qquad \text{s.t.} \quad \sum_{i=0}^{m} \gamma_i=1.
$$

The backward pass uses equilibrium differentiation:
$$
\frac{\partial \alpha^*}{\partial \Theta } =
\left ( \mathbf{I} - \frac{\partial f_{\Theta}(\alpha^*, z)}{\partial \alpha^*} \right )^{-1}
\frac{\partial f_{\Theta}(\alpha^*, z)}{\partial \Theta}.
$$
Because direct inversion is expensive, the implementation uses **phantom gradients** based on a truncated Neumann series,
$$
\frac{\partial \alpha^*}{\partial \Theta} =
\sum_{l=0}^{L-1} \left ( \frac{\partial f_{\Theta}}{\partial \alpha^*} \right )^l
\frac{\partial f_{\Theta}}{\partial \Theta},
$$
with truncation length $L=5$ [2508.15553].

The equilibrium-training machinery places DECSC in continuity with broader DEQ methodology for inverse problems and with earlier equilibrium sparse coding, but its equilibrium variable retains a true CSC interpretation rather than collapsing sparse inference into generic latent-state refinement [2201.06931], [2203.15901].

## 5. Architectural realization for hyperspectral images

The architectural novelty of DECSC lies in the way HSI-specific priors are embedded directly into the equilibrium map rather than appended as post-processing modules.

The **GIC branch** uses shared 2D CSC followed by a **Swin Transformer**. Since the shared code $S$ represents globally common spatial structure, the transformer is inserted after the 2D sparse coding update to model **nonlocal spatial self-similarity**. The attention mechanism is described by
$$
Q_i = W_i^q P,\qquad K_i = W_i^k P,\qquad V_i = W_i^v P,
$$
$$
\text{head}_i = \text{Softmax}\left( \frac{(W^q_iP)(W^k_iP)^{T}}{\sqrt{d_i}} \right) W^v_iP,
$$
and
$$
\text{MSA}(P) = \text{Concat}(\text{head}_1,\ldots,\text{head}_h)W^o.
$$
Implementation details reported for this branch are a dictionary size of $M=192$, 2D filter size $9\times 9$, a Swin Transformer with four stacked stages, and window size $4\times 4$ [2508.15553].

The **LSU branch** uses unshared 3D CSC followed by a **detail enhancement module**. Its purpose is to preserve local spatial-spectral correlations and high-frequency structures that may otherwise be smoothed away by denoising. The branch uses dictionary size $J=96$ and 3D filter size $9\times 9\times 3$ [2508.15553].

The detail enhancement module is built from **difference convolution (DConv)** and **spatial attention** implemented by cascaded 3D convolutions:
$$
\begin{split}
S &= S + \text{DConv}(S), \\
S &= S + S \bigodot \text{Conv}_1(\text{Conv}_1(S)).
\end{split}
$$
The DConv operator is extended for HSI using one standard 3D convolution and four HSI-specific edge extraction operators: **central difference**, **inter-band difference**, **horizontal difference**, and **vertical difference**. The resulting features are then processed by two cascaded 3D convolutions acting as an attention mechanism that adaptively emphasizes detail-rich regions [2508.15553].

This architecture expresses a strong inductive decomposition: shared 2D CSC for global spatial consistency across bands, transformer-based modeling for nonlocal self-similarity, unshared 3D CSC for local spatial-spectral detail, and a detail-focused regularizer for edge preservation. A plausible implication is that DECSC is especially well matched to HSI denoising scenarios in which both band-consistent structure and band-specific fine detail must be retained simultaneously.

## 6. Optimization, datasets, and empirical profile

Training uses the Euclidean reconstruction objective
$$
\mathcal{L}_{\Theta}=\frac{1}{N}\sum_{i}^{N} \left\| \widehat{X}_i-X_i\right\|_F^2.
$$
The implementation is in **PyTorch**, with **Adam**, initial learning rate $10^{-4}$, batch size $8$, learning rate halved every 10 epochs, and 30 training epochs. The model is pre-trained on **ICVL** for synthetic noise removal and then fine-tuned on the corresponding test datasets [2508.15553].

The synthetic experiments use ICVL and Houston 2018. Real-noise cases include **EO-1** and **GF-5 CapitalAirport**, where evaluation is qualitative because ground truth is unavailable. Metrics are **PSNR**, **SSIM**, and **SAM**. The baselines span model-driven, data-driven, and hybrid-driven methods, including BM4D, MTSNMF, LLRT, NGMeet, LRMR, E-3DTV, 3DlogTNN, SST, TRQ3D, SERT, T3SC, and MTSNN++ [2508.15553].

| Setting | DECSC result (PSNR / SSIM / SAM) | Note |
|---|---|---|
| ICVL, non-i.i.d. Gaussian $[0,95]$ | **45.64 / 0.9848 / 0.0387** | Better than all listed methods |
| ICVL, mixture noise | **42.67 / 0.9756 / 0.0625** | Better than all listed methods |
| ICVL, correlated-variance noise | **48.06 / 0.9891 / 0.0298** | SERT has higher PSNR/SSIM |
| Houston 2018, non-i.i.d. Gaussian $[0,95]$ | **43.25 / 0.9834 / 0.0305** | Best values in the table |
| Houston 2018, mixture noise | **38.65 / 0.9581 / 0.0480** | Best values in the table |
| Houston 2018, correlated-variance noise | **46.32 / 0.9904 / 0.0225** | SERT is stronger |

The ablation study isolates the Swin Transformer, difference convolution, and attention mechanism. On ICVL under $[0,95]$ Gaussian noise, the full model achieves **45.64 / 0.9848 / 0.0387**; removing Swin yields **44.87 / 0.9823 / 0.0431**; removing DConv yields **45.32 / 0.9839 / 0.0410**; and removing attention yields **44.99 / 0.9829 / 0.0434** [2508.15553]. The dictionary size study reports best performance at **GIC atoms = 192** and **LSU atoms = 96**, while the Neumann truncation study reports best performance at **$L=5$**.

A convergence study shows PSNR improving and then stabilizing as iterations increase, which supports the practical equilibrium interpretation. At the same time, the computational cost is substantial: for the Swin-based DECSC, the reported profile is **4.29M** parameters, **36.95 s** inference time, and **56.39T** FLOPs. The paper notes a Mamba-based variant with comparable performance and reduced runtime [2508.15553].

## 7. Interpretation, limitations, and recurring points of confusion

DECSC is best understood as an explicit CSC model whose iterative solver has been transformed into a DEQ. It is therefore neither a generic deep denoiser nor a generic equilibrium inverse-problem network. The sparse variables, convolutional dictionaries, proximal soft-thresholding, and synthesis reconstruction remain explicit throughout the formulation [2508.15553].

Several neighboring methods are often conflated with DECSC but should be distinguished. DEQ-based imaging models such as the video SCI formulation and the ISTA-inspired compressive sensing reconstruction network operate with equilibrium maps over reconstructed signals or denoised states, not explicit convolutional sparse codes [2201.06931], [2401.02884]. Deep equilibrium sparse coding with a matrix dictionary is closer in spirit, but it remains blockwise and non-convolutional in the sparse coding operator [2203.15901]. Learned convolutional sparse coding supplies the correct convolutional synthesis and sparse-inference structure, but in finite unrolled form rather than equilibrium form [1711.00328].

The main reported strengths of DECSC are its physically structured HSI prior, clearer optimization-network alignment than finite unfolding, interpretable module decomposition, and strong performance under non-i.i.d. Gaussian and mixture noise [2508.15553]. Its principal limitations are equally explicit: high computational cost, limited formal convergence theory specialized to the DECSC operator, possible weakness on some globally correlated spectral patterns where SERT is stronger, and sensitivity to solver settings, truncation length, and Jacobian behavior [2508.15553].

A further misconception is that DEQ automatically supplies a rigorous convergence guarantee for any learned equilibrium map. The DECSC paper states that the architecture “seamlessly integrates with numerical solvers to reach a stable equilibrium and therefore has convergence guarantee,” yet the provided formulation does not include an explicit contraction or monotonicity theorem specialized to the DECSC update map. This suggests that the method is theoretically motivated by DEQ principles and empirically supported by stabilizing PSNR-versus-iteration behavior, but not accompanied by a full task-specific convergence proof.

In the broader history of sparse modeling, DECSC can therefore be read as the point where three lines of work meet: the canonical CSC objective and its proximal/alternating solvers, learned tied-weight convolutional sparse inference, and DEQ fixed-point computation with implicit training. Within that synthesis, its most distinctive contribution is not merely the use of equilibrium machinery, but the use of equilibrium machinery to preserve an explicit convolutional sparse coding interpretation while incorporating HSI-specific nonlocal and detail-aware priors.

Source: https://www.emergentmind.com/topics/deep-equilibrium-convolutional-sparse-coding-decsc