Convolutional Dictionary Learning

• Convolutional Dictionary Learning is a framework that represents signals as convolutions of learned filters with sparse activation maps, ensuring translation invariance.
• It alternates between convolutional sparse coding and dictionary update steps using methods like ISTA, ADMM, and unrolled networks to balance efficiency and interpretability.
• Its wide applications in image denoising, MRI, neuroscience, and astronomy demonstrate its versatility in handling large-scale, complex, and heterogeneous data.

Convolutional Dictionary Learning (CDL) is a sparse representation framework in which a signal or image is modeled as a superposition of convolutions between learned local filters and sparse activation maps. In contrast to patch-based dictionary learning, CDL builds translation invariance directly into the model by using convolution over the full signal domain, so that a single atom represents all of its shifts. Across the literature, CDL appears both as a classical inverse-problem formulation—typically alternating between convolutional sparse coding and dictionary update—and as a foundation for online, distributed, probabilistic, and unrolled neural architectures. The common synthesis-model perspective is that recurring local structure is captured by a small bank of filters, while sparsity in the coefficient maps encodes when and where those structures occur (Garcia-Cardona et al., 2017, Janjušević et al., 2021).

1. Mathematical formulation and representational assumptions

The canonical single-signal convolutional sparse coding problem seeks sparse feature maps for a fixed convolutional dictionary: $$\min_{x}\; \frac{1}{2}\left\| y - \sum_{k=1}^K d_k * x_k \right\|_2^2 + \lambda \|x\|_1,$$ where $y$ is the observed signal, $\{d_k\}$ are filters, $\{x_k\}$ are coefficient maps, and $*$ denotes convolution. In operator form, if $D$ denotes the convolutional dictionary, the same problem is

$$\min_x\; \frac{1}{2}\|y - D x\|_2^2 + \lambda \|x\|_1.$$

Over a dataset $\{y_i\}$, CDL jointly optimizes filters and codes under norm constraints to remove scale ambiguity: $$\min_{\{x_i\},D}\; \sum_i \left[\frac{1}{2}\left\| y_i - \sum_k d_k * x_{i,k}\right\|_2^2 + \lambda \|x_i\|_1\right] \quad \text{s.t.} \quad \|d_k\|_2 \le 1.$$ In this synthesis view, the transpose $D^\top$ corresponds to cross-correlation; channelwise, $y$0, where $y$1 is the spatially flipped filter (Janjušević et al., 2021).

A recurrent misconception is that CDL is merely patch-based dictionary learning with convolution added afterward. The technical distinction is stronger: patch methods model independent local vectors, whereas CDL couples all shifts through a block-convolutional operator, yielding global translation invariance at full-image or full-signal scale. This changes both the statistical efficiency and the computational structure of the problem, because the same atom explains occurrences at arbitrary positions without duplicating shifted copies in the dictionary (Garcia-Cardona et al., 2017).

The model admits several important variants. Multi-image and multi-channel formulations share coefficient maps across channels while assigning channel-specific filters. Masked formulations replace the data-fidelity term by $y$2 to handle missing data or suppress FFT-induced boundary artifacts. Boundary conditions themselves vary across the literature: some methods assume circular convolution to exploit FFT diagonalization, others adopt zero-padding, and others use explicit masks or truncation operators to retain valid-region fidelity (Garcia-Cardona et al., 2017, Chun et al., 2017, Moreau et al., 2019).

2. Optimization algorithms and identifiability structure

Most CDL algorithms alternate between two convex subproblems: convolutional sparse coding with fixed filters, and dictionary update with fixed sparse maps. For the sparse-coding block, proximal methods dominate. The $y$3 term induces soft-thresholding through

$y$4

and a standard ISTA step for fixed $y$5 is

$y$6

with $y$7, $y$8 the Lipschitz constant of $y$9. FISTA adds momentum; ADMM introduces auxiliary variables and dual updates; coordinate-descent methods exploit the locality of convolutional interference neighborhoods (Janjušević et al., 2021, Garcia-Cardona et al., 2017).

The dictionary-update block is more delicate. In the review literature, a central observation is that CSC and CDL are not symmetric in difficulty: the CSC normal equations decompose into small rank-1 systems in the DFT domain, whereas dictionary update involves rank-$\{d_k\}$0 components per frequency, making the linear algebra substantially harder. Comparative results identify two especially effective families in 2017-era batch CDL: FISTA-based dictionary updates in serial settings, and ADMM consensus in parallel settings. The same review also reports that coupling alternating blocks through ADMM auxiliary variables, rather than primary variables, is markedly more stable and supports one-iteration alternation (Garcia-Cardona et al., 2017).

Majorization-based methods were proposed partly to avoid the parameter sensitivity of augmented-Lagrangian schemes. The Block Proximal Gradient using a Majorizer (BPG-M) framework constructs block-separable quadratic majorizers of the CDL Hessian, yielding monotone or near-monotone descent, convergence of limit points to Nash points, and substantially lower memory demand than state-of-the-art ADMM variants in single-threaded large-scale settings. In image denoising experiments, filters learned by BPG-M-based CDL outperform filters trained by ADMM under relatively strong AWGN (Chun et al., 2017).

A different line of work recasts CDL as structured tensor factorization. Under a convolutional ICA model with independent latent coordinates, third-order cumulants of the observations admit a CP decomposition whose factor matrices are constrained to be stacked circulant. The resulting constrained CP-ALS algorithm projects unconstrained least-squares updates onto the space of circulant matrices via FFT-domain diagonal extraction. This approach is embarrassingly parallel, uses FFTs and matrix multiplications as primitive operations, and empirically converges to the dictionary faster and more accurately than alternating minimization over filters and activations. Its guarantees are computational and stationary-point based rather than global-optimality guarantees (Huang et al., 2015).

3. Scaling CDL to large datasets and continuous shifts

A persistent difficulty in CDL is that batch formulations scale poorly with dataset size because they require simultaneous access to all training samples or all sparse codes. Online CDL addresses this by replacing the full empirical objective with a surrogate summarized by sufficient statistics. In one formulation, the dictionary update is driven by frequency-domain accumulators $\{d_k\}$1 and $\{d_k\}$2, so memory scales as $\{d_k\}$3 rather than the $\{d_k\}$4 cost of a direct extension to the convolutional setting. A forgetting schedule $\{d_k\}$5 biases the surrogate toward recent data while asymptotically recovering the unweighted average (Liu et al., 2017).

A related online line develops first- and second-order stochastic methods. The projected-SGD variant updates the dictionary using the current sample only, whereas the second-order method accumulates curvature terms $\{d_k\}$6 and gradient terms $\{d_k\}$7, then solves a constrained quadratic subproblem by FISTA. With diminishing fixed-point residual tolerances, the paper proves $\{d_k\}$8 and convergence to the set of stationary points almost surely. The same framework extends to masked data by inserting the observation mask directly into the gradient and accumulators (Liu et al., 2017).

Approximate online CDL further reduces cost by introducing per-sample auxiliary dictionaries that decouple cross-filter couplings in the dictionary update. The resulting method brings online memory from $\{d_k\}$9 down to $\{x_k\}$0, with overall complexity $\{x_k\}$1. Extensive evaluations on image datasets show that this approximation substantially lowers computation while preserving the effectiveness of state-of-the-art OCDL algorithms, and it remains feasible for dictionary sizes at which earlier online methods become memory-intractable (Veshki et al., 2023).

Distributed CDL addresses a different scaling axis: extremely large images and signals. DiCoDiLe partitions the domain across workers, uses Locally Greedy Coordinate Descent for sparse coding, and enforces asynchronous consistency through a soft-lock mechanism that suppresses conflicting updates without a central server. The dictionary update depends on distributed sufficient statistics $\{x_k\}$2 and $\{x_k\}$3, making its cost independent of the full signal size $\{x_k\}$4. The method scales to a $\{x_k\}$5 Hubble Space Telescope image with $\{x_k\}$6, $\{x_k\}$7 atoms, and $\{x_k\}$8 workers, where it learns interpretable atoms capturing stars and localized astronomical structures (Moreau et al., 2019).

Classical CSC also suffers from time-quantization when the underlying events are continuous-time but observed on a discrete grid. Off-grid CDL remedies this by expanding the dictionary with interpolated atoms generated by bandlimited sinc shifts. The resulting COMP-INTERP sparse coder achieves CBP-like localization accuracy while being two orders of magnitude faster on simulated data, and the accompanying off-grid dictionary update yields more accurate templates than standard on-grid updates (Song et al., 2019).

4. Unrolled, interpretable, and physics-informed neural instantiations

A major recent development is the reinterpretation of CDL solvers as trainable iterative networks. In CDLNet, a $\{x_k\}$9-layer network unrolls an ISTA-style sparse-coding procedure with untied layer parameters: $*$0 Here $*$1 is a learned analysis operator, $*$2 a learned synthesis operator, $*$3 a final synthesis dictionary, and $*$4 nonnegative subband-wise thresholds. The architecture deliberately avoids batch normalization, residual learning, and feature-domain $*$5 convolutions; every layer corresponds directly to a sparse-coding iteration (Janjušević et al., 2021).

This interpretability does not imply weak empirical performance. On BSD68 with matched train/test noise, Big-CDLNet-S reaches PSNRs of $*$6, $*$7, and $*$8 dB at $*$9, respectively, compared with $D$0, $D$1, and $D$2 for DnCNN at similar parameter count. In wide-range blind denoising with training on $D$3, Big-CDLNet-A attains $D$4 dB at $D$5, versus $D$6 for BF-DnCNN and $D$7 for DnCNN-B, illustrating that explicit noise-adaptive thresholding can produce near-perfect generalization to unseen noise levels (Janjušević et al., 2021).

The same design extends to joint denoising and demosaicing. In the JDD variant, the mask $D$8 is inserted into the data-fidelity gradient so that only CFA-observed entries contribute to the residual. Thresholds are parameterized as affine functions of the noise level, $D$9, preserving a direct MAP-style relation between the observation noise and shrinkage magnitude. This model achieves competitive or superior denoising and JDD performance at comparable parameter counts in both grayscale and color settings, and the same unrolled template supports MC-SURE-based unsupervised training (Janjušević et al., 2021).

Unrolled CDL has also been generalized beyond Gaussian image restoration. For natural exponential-family data, convolutional dictionary learning can be mapped to unfolded constrained auto-encoders whose encoder uses the distribution-specific inverse link $$\min_x\; \frac{1}{2}\|y - D x\|_2^2 + \lambda \|x\|_1.$$0 inside the iterative updates. The resulting architectures support Poisson and binomial observations, competitive supervised Poisson denoising, and unsupervised latent-feature discovery in neural spiking data (Tolooshams et al., 2019). In dynamic MRI, a physics-informed iterative neural network embeds CDL inside an $$\min_x\; \frac{1}{2}\|y - D x\|_2^2 + \lambda \|x\|_1.$$1 data-consistency solve, learns both the filters and the regularization parameters end-to-end, and improves over conventional model-agnostic training while remaining fully interpretable (Kofler et al., 2022). In hybrid-field XL-RIS channel estimation, CDL is similarly cast as a bilevel problem and unrolled into Convolutional ISTA-Net, CISTA-Net+, and CNN-CDL; with $$\min_x\; \frac{1}{2}\|y - D x\|_2^2 + \lambda \|x\|_1.$$2 pilots at $$\min_x\; \frac{1}{2}\|y - D x\|_2^2 + \lambda \|x\|_1.$$3 dB SNR, CNN-CDL attains NMSE $$\min_x\; \frac{1}{2}\|y - D x\|_2^2 + \lambda \|x\|_1.$$4 dB, only $$\min_x\; \frac{1}{2}\|y - D x\|_2^2 + \lambda \|x\|_1.$$5 dB from the OLS oracle (Zheng et al., 2024).

5. Structured priors, weak supervision, and personalized CDL

One limitation of standard CDL is that it treats dictionary atoms as largely unstructured beyond support and norm constraints. Gaussian Process Convolutional Dictionary Learning introduces explicit GP priors on templates. In the Gaussian, stationary-kernel, non-overlapping setting, the GP-regularized dictionary update is equivalent to applying a Wiener filter to the unregularized template estimate in the frequency domain: $$\min_x\; \frac{1}{2}\|y - D x\|_2^2 + \lambda \|x\|_1.$$6 where $$\min_x\; \frac{1}{2}\|y - D x\|_2^2 + \lambda \|x\|_1.$$7 is the kernel PSD. This suppresses high-frequency components in low-data or low-SNR regimes and produces smoother, more interpretable templates than unregularized CDL (Song et al., 2021).

Weak supervision has motivated another extension. In a multi-instance multi-label setting, weakly supervised CDL decomposes each bag-level signal into a shared background dictionary $$\min_x\; \frac{1}{2}\|y - D x\|_2^2 + \lambda \|x\|_1.$$8 and class-specific dictionaries $$\min_x\; \frac{1}{2}\|y - D x\|_2^2 + \lambda \|x\|_1.$$9, with a nuclear-norm penalty on $\{y_i\}$0 to prevent feature dilution. Optimization uses a Block Proximal Gradient method with Majorization, while a learnable pooling operator aggregates instance-level activations into bag-level labels. On ESC-10, the reported five-run average gives accuracy $\{y_i\}$1 and AUC $\{y_i\}$2, outperforming the listed MIML baselines in low-label regimes (Chen et al., 11 Mar 2025).

Population heterogeneity motivates yet another generalization. Personalized CDL models each subject-specific local dictionary as a learnable transformation of a global dictionary, with time warping and optional rotations as admissible transformations. In the balanced case $\{y_i\}$3, the paper proves a mixed-effects-style rate $\{y_i\}$4 for estimating the global atom under its assumptions. Empirically, on gait analysis the method attains sensitivity $\{y_i\}$5 versus $\{y_i\}$6 for the state-of-the-art template-matching baseline, while preserving a coherent global gait-cycle atom and exposing clinically meaningful deviations in orthopedic and neurological groups. On PTB-XL ECG data, an SVM on the learned personalization parameters reaches macro one-vs-one AUC $\{y_i\}$7 (Roques et al., 10 Mar 2025).

These structured variants share a common implication: CDL is not restricted to learning unconstrained shift-invariant filters. It can also encode smoothness, background-versus-discriminative separation, and subject-specific deformation structure without abandoning the synthesis-model and sparse-coding foundations.

6. Applications, empirical behavior, and open issues

CDL has been applied across a wide range of domains. In image processing it underlies denoising, inpainting, demosaicing, and dynamic MRI reconstruction; in neuroscience it supports spike sorting and latent stimulus recovery; in astronomy it discovers recurring structures in massive telescope images; in physiology it models gait and ECG cycles; in wireless communications it estimates hybrid-field channels; and in rare-event analysis it has been reoriented toward anomaly discovery (Janjušević et al., 2021, Janjušević et al., 2021, Moreau et al., 2019, Kofler et al., 2022, Zheng et al., 2024).

Spike sorting is an especially clear example of CDL’s fit to domain structure. Extracellular recordings are modeled as convolutions between neuron-specific waveforms and sparse spike trains, with refractoriness and limited simultaneity providing additional structure. The cOMP/cKSVD pipeline developed for this setting yields theoretical sample-complexity bounds as a function of neuron count, overlap rate, and firing rate, and on tetrode data it achieves true miss below $\{y_i\}$8, outperforming the paper’s K-means baseline (Song et al., 2018).

Robustness to outliers and rare phenomena has recently become explicit. RoseCDL combines stochastic windowing with inline outlier detection based on patch-level reconstruction errors, trimming high-error windows during learning so that common structure is learned first and rare events remain isolated in the residual. On signals up to $\{y_i\}$9 samples, RoseCDL runs in $$\min_{\{x_i\},D}\; \sum_i \left[\frac{1}{2}\left\| y_i - \sum_k d_k * x_{i,k}\right\|_2^2 + \lambda \|x_i\|_1\right] \quad \text{s.t.} \quad \|d_k\|_2 \le 1.$$0 s whereas the $$\min_{\{x_i\},D}\; \sum_i \left[\frac{1}{2}\left\| y_i - \sum_k d_k * x_{i,k}\right\|_2^2 + \lambda \|x_i\|_1\right] \quad \text{s.t.} \quad \|d_k\|_2 \le 1.$$1CSC baseline is not reported beyond $$\min_{\{x_i\},D}\; \sum_i \left[\frac{1}{2}\left\| y_i - \sum_k d_k * x_{i,k}\right\|_2^2 + \lambda \|x_i\|_1\right] \quad \text{s.t.} \quad \|d_k\|_2 \le 1.$$2 samples. On the TSB-UAD benchmark summary, RoseCDL reports AUC ROC $$\min_{\{x_i\},D}\; \sum_i \left[\frac{1}{2}\left\| y_i - \sum_k d_k * x_{i,k}\right\|_2^2 + \lambda \|x_i\|_1\right] \quad \text{s.t.} \quad \|d_k\|_2 \le 1.$$3 on Simulated data and $$\min_{\{x_i\},D}\; \sum_i \left[\frac{1}{2}\left\| y_i - \sum_k d_k * x_{i,k}\right\|_2^2 + \lambda \|x_i\|_1\right] \quad \text{s.t.} \quad \|d_k\|_2 \le 1.$$4 on ECG, with average runtime $$\min_{\{x_i\},D}\; \sum_i \left[\frac{1}{2}\left\| y_i - \sum_k d_k * x_{i,k}\right\|_2^2 + \lambda \|x_i\|_1\right] \quad \text{s.t.} \quad \|d_k\|_2 \le 1.$$5 s versus $$\min_{\{x_i\},D}\; \sum_i \left[\frac{1}{2}\left\| y_i - \sum_k d_k * x_{i,k}\right\|_2^2 + \lambda \|x_i\|_1\right] \quad \text{s.t.} \quad \|d_k\|_2 \le 1.$$6 s for AE, $$\min_{\{x_i\},D}\; \sum_i \left[\frac{1}{2}\left\| y_i - \sum_k d_k * x_{i,k}\right\|_2^2 + \lambda \|x_i\|_1\right] \quad \text{s.t.} \quad \|d_k\|_2 \le 1.$$7 s for MP, and $$\min_{\{x_i\},D}\; \sum_i \left[\frac{1}{2}\left\| y_i - \sum_k d_k * x_{i,k}\right\|_2^2 + \lambda \|x_i\|_1\right] \quad \text{s.t.} \quad \|d_k\|_2 \le 1.$$8 s for OC-SVM (Yehya et al., 9 Sep 2025).

Despite its breadth, CDL retains several unresolved issues. The global problem is nonconvex; convergence claims are typically limited to stationary points, Nash points, or monotone decrease of constrained surrogates rather than global optimality. Performance remains sensitive to boundary handling, coherence, overcompleteness, and hyperparameter selection in classical formulations. Tensor and personalized variants likewise do not provide general global guarantees for the full alternating scheme. At the same time, the unrolled-network literature shows that many of these parameters can be absorbed into learned operators and thresholds, suggesting a continuing synthesis between classical sparse modeling and task-specific end-to-end training (Chun et al., 2017, Huang et al., 2015, Roques et al., 10 Mar 2025).

Taken together, the literature presents CDL as a family of models rather than a single algorithm. Its invariant core is the convolutional synthesis prior with sparse activations; its diversity lies in how sparsity is enforced, how dictionaries are updated, how scale is managed, how physics or supervision is introduced, and how learning is stabilized under scale, noise, heterogeneity, or distributional mismatch.