Abstract
Hyperspectral image (HSI) deconvolution is a challenging ill-posed inverse problem, made difficult by the data's high dimensionality. We propose a parameter-parsimonious framework based on a low-rank Canonical Polyadic Decomposition (CPD) of the entire latent HSI $\mathbf{\mathcal{X}} \in \mathbb{R}^{P\times Q \times N}$. This approach recasts the problem from recovering a large-scale image with $PQN$ variables to estimating the CPD factors with $(P+Q+N)R$ variables. This model also enables a structure-aware, anisotropic Total Variation (TV) regularization applied only to the spatial factors, preserving the smooth spectral signatures. An efficient algorithm based on the Proximal Alternating Linearized Minimization (PALM) framework is developed to solve the resulting non-convex optimization problem. Experiments confirm the model's efficiency, showing a numerous parameter reduction of over two orders of magnitude and a compelling trade-off between model compactness and reconstruction accuracy.
Introduction
Hyperspectral images (HSIs) are 3D "data cubes" that capture information across dozens or hundreds of spectral bands, providing a much richer view of a scene than standard RGB images. However, during acquisition, these images often suffer from blurring (convolution) and noise. The task of deconvolution—recovering the original clean image—is notoriously difficult because of the "curse of dimensionality."
A typical HSI might have millions of variables. Existing solutions generally fall into two camps: full-rank models that operate on the entire image at once (accurate but computationally heavy) and patch-based models that break the image into small 3D chunks (efficient but often losing global structure). This paper proposes a middle ground: a method that is computationally efficient, respects global structure, and uses as few parameters as possible.
Methodology
The proposed framework is built upon three pillars designed to exploit the inherent structure of HSI data:
- Canonical Polyadic Decomposition (CPD): Instead of treating the HSI as a giant block of $P \times Q \times N$ pixels, the model decomposes it into three factor matrices ($A, B, C$). $A$ and $B$ represent spatial features, while $C$ represents spectral signatures. This reduces the number of variables from $PQN$ to $(P+Q+N)R$, where $R$ is the "rank" of the tensor.
- Anisotropic Total Variation (TV) Regularization: Spatial dimensions usually contain sharp edges (objects), while spectral signatures are typically smooth. The authors apply TV regularization—which encourages piecewise constant patterns—only to the spatial factors ($A, B$). This preserves sharp object boundaries without introducing artifacts into the smooth spectral data.
- PALM Algorithm: To solve the resulting non-convex optimization problem, the researchers utilize the Proximal Alternating Linearized Minimization (PALM) framework, accelerated by Discrete Fourier Transforms (DFT) to handle 2D convolutions efficiently.
A key insight of this approach is the shift from image recovery to parameter estimation. By solving for the CPD factors rather than the pixels themselves, the optimization landscape becomes significantly smaller. As shown in the figure above, reconstruction quality (PSNR) peaks at a specific rank ($R \approx 20$) and then plateaus, confirming that HSIs are naturally low-rank and that excessive parameters eventually overfit noise.
Experimental Evaluation
The method was evaluated using the Columbia Multispectral Database (CAVE), containing 32 HSIs sized $512 \times 512 \times 31$. To simulate deconvolution, the authors applied a $9 \times 9$ Gaussian blur and added white Gaussian noise to the clean images.
Key Results
The proposed method demonstrates a significant leap in efficiency compared to full-rank methods like 3DFTV or WLRTR.
- Parameter Reduction: While competing methods require roughly 8 million parameters, the proposed model requires only 30,000—a reduction of over two orders of magnitude.
- Memory Efficiency: Memory usage dropped from 62 MB to just 0.24 MB.
- Performance Trade-off: While the proposed method has a slightly higher error rate (RMSE of 6.99 vs. ~4.4 for full-rank methods), it provides a compelling trade-off for resource-constrained environments like satellites or mobile sensors.
Discussion and Conclusion
The proposed framework achieves extreme compactness by recasting HSI deconvolution as a low-rank parameter estimation problem. While there is a noticeable gap in reconstruction accuracy compared to full-rank models, the 200x reduction in parameters and memory usage makes it a viable candidate for edge computing. Future work will focus on automating the selection of the optimal rank $R$ and improving noise robustness within the CPD framework to bridge the accuracy gap.
Summary updated