Hyperspectral imaging (HSI) captures measurements at multiple wavelengths across the electromagnetic spectrum, providing information that improves material segmentation and classification beyond RGB imagery. While HSI devices often acquire a large number of spectral bands, this increases both cost and acquisition time. However, not all bands contribute equally to task-specific performance. We propose a deep spectral band selection (DSBS) framework for HSI tasks. Unlike methods that preserve non-task-specific information, DSBS identifies the most informative bands for a given task by jointly training a fully differentiable band selector and a neural network in an end-to-end (E2E) learning scheme. The selection is guided by a bin function and an 1p-norm regularizer to reach a target number of bands. Experiments on segmentation and classification show that DSBS outperforms state-of-the-art machine learning and deep learning methods. Results show that DSBS outperforms state-of-the-art E2E methods for HSI classification by about 8% across the evaluated metrics and yields an improvement of 1% over using all spectral bands for material segmentation. Additionally, we validate DSBS in a testbed implementation, starting from full-spectrum images (301 bands). The end-to-end training converges to 10 wavelengths; under crossvalidation, considering only these 10 bands yields competitive cocoa-bean classification, with an overall accuracy of 76.40%, retaining approximately 95.5% of the accuracy observed in the 10-band end-to-end evaluation (80.01%) while reducing acquisition speed by a factor of 30 times. (c) 2026 Optica Publishing Group. All rights, including for text and data mining (TDM), Artificial Intelligence (AI)training, and similar technologies, are reserved.

Self-supervised learning for imaging inverse problems is increasingly important in photon-limited settings, where acquiring clean ground truth is impractical and reconstruction must remain stable under dataset and acquisition shifts. This challenge is amplified under Poisson noise, whose signal-dependent statistics interact with sampling operators (e.g., CFA mosaicing). Meanwhile, foundation vision encoders trained at web scale offer distortion-invariant, content-related representations that generalize well across domains, suggesting a promising route to build priors that transfer beyond the training distribution without expensive fine-tuning. This paper proposes an ADMM-inspired unrolled plug-and-play solver for Poisson inverse problems that decouples a closed-form data-consistency update from a parameter-efficient prior. The prior is implemented as a lightweight decoder operating on frozen CLIP RN50 dense multi-scale features, adapting foundation representations with less trainable parameters. For self-supervision, the method integrates GR2R measurement-domain re-corruption with an Equivariant Imaging regularizer via virtual acquisitions. Experiments on Poisson CFA demosaicing and deblurring show competitive quality, improved robustness under shifts, and self-supervised performance approaching supervised training.

Compressive imaging (CI) consists of reconstructing images from incomplete observed data. The reconstruction process involves solving an ill-posed inverse problem which is highly dependent on the number of real measurements, with a greater number of measurements typically leading to more accurate reconstructions. Due to their ability to learn data distributions, diffusion models (DM) have emerged as promising techniques for various inverse problems. Mainly, DMs solve inverse problems by conditioning the generation process to the acquired measurements. In this work, we introduce a new approach to improve this conditioning by exploiting synthetic measurements, which come from a synthetic sensing matrix. Synthetic measurements are estimated from real data via a neural network. The combined real and synthetic measurements form an augmented set, which is input into the conditional DM to enhance reconstruction capacity. Computational experiments demonstrate that augmenting measurements with the conditional DM improves performance compared to using only real measurements.

Computational Imaging (CI) enables the acquisition of low-dimensional or corrupted observations that can be recov-ered as high-quality images by solving an inverse problem using computational algorithms. The Alternating Direction Method of Multipliers (ADMM) is widely used due to its robustness in han-dling constraints and regularization techniques. Recent advances in generative models have established them as powerful priors that can be integrated into ADMM to improve image recovery in challenging ill-posed inverse problems. However, ADMM with generative priors typically starts the recovery process by sampling the latent space from a standard normal distribution, particularly in highly underdetermined settings. This random initialization causes the recovery to begin at arbitrary points in the output space of the generator, hindering convergence and reducing the likelihood of reaching optimal solutions. Therefore, a two-stage approach is proposed for learning the latent space of the generative model to improve recovery performance. In the first stage, the latent space is estimated from the observations using an encoder model that leverages a pre-trained generative model. In the second stage, the learned latent space is used to initialize the ADMM algorithm with the generator model serving as the prior. Experimental results show that the proposed method improves PSNR by approximately 2 dB in the most challenging cases, outperforming the ADMM algorithm with standard normal latent space initialization.

Signal reconstruction from compressed, noisy observations is a ubiquitous challenge in various applications. To address its ill-posed nature, a suitable prior of the underlying signal is required. Generative adversarial networks (GANs) emerge as a natural prior, enabling realistic reconstructions. However, existing approaches either optimize a GAN conditioned on the measurements from scratch or use pre-trained GANs to find images that best fit real measurements. We propose an alternative GAN-based method that, instead of sampling directly from the signal distribution, generates low-dimensional synthetic observations from the real ones. An adversarial self-distillation strategy optimizes the GAN, extracting meaningful signal information for synthetic measurement generation. These samples form an augmented measurement set, improving the conditioning of compressed sensing solvers, including model-based and deep learning-based methods. We validate our approach on MNIST with a binary sensing matrix for the single-pixel camera, achieving significant improvements in ADMM, PnP-ADMM, and Unrolled ADMM using generated measurements.

In remote sensing and environmental monitoring, the authenticity of the collected hyperspectral image (HSI) is critical since unauthorized changes could cause inaccurate evaluation. Traditional approaches ensure HSI integrity after the acquisition, leaving the data vulnerable to unauthorized access or modification before applying a protective strategy. In response, this work presents a pushbroom HSI system that integrates an authentication key on the acquisition process, ensuring the key imperceptibility without affecting performance in spectral classification tasks. In addition, this work develops an algorithm designed to simulate attacks through class modifications in the HSI. Additionally, a detection algorithm is proposed to identify modifications by comparing the signed HSI with the altered version, effectively recognizing the embedded key. The proposed method was tested using simulated data and acquired data in the optical laboratory. Results demonstrate that the system maintains classification accuracy, further enhancing imperceptibility. Additionally, modifications are detectable by the key with 99% accuracy.

In compressive imaging systems, the scene is acquired via linear coded noisy projections, known as measurements, requiring a recovery process to estimate the underlying signal. This recovery is inherently ill-posed, posing a challenge for accurate signal recovery. Existing methods that employ prior information about the signal often fail in practical scenarios. In this work, instead of developing a new prior over the signal, we exploit the structure of the low-dimensional measurements to synthesize an augmented measurement set that can be used in various recovery methods to improve its performance. We used a deep neural network to generate the synthetic measurements from the acquired data. We show the benefits of this approach in two schemes, deep learning-based recovery and the plug-and-play (PnP) algorithm. Particularly, our method is interpreted as a non-linear preconditioning technique for the PnP algorithm. We show improved performance for different sensing matrices.

Spectral imaging is a technique that enables the acquisition and analysis of radiation emitted by incident light from a scene. Spectral acquisition involves various scanning strategies, such as snapshot spectral image acquisition based on compressive sensing theory. However, scanning methods require meticulous calibration processes in the optical setups and pose challenges in implementation, especially in uncontrolled environments. Current research has employed generative adversarial networks (GANs) to produce new spectral images and alleviate reliance on complex optical setups. In addition, traditional methods are focused on RGB-to-spectral mapping techniques, where new spectral images are not created. Therefore, this work proposes spectral imaging generation through RGB image dataset guidance by using GANs. Specifically, a generative model can produce spectral images from random noise input and map them to RGB images through a spectral response matrix, which is fed into a discriminator model for adversarial training. To ensure realistic spectral image generation, an implicit learning approach to spectral information is introduced, where a pretrained model with spectral images is used to regularize the generated spectral images during training. Finally, a post-processing step normalizes the mean and standard deviation of generated spectral images according to each spectral band of the real training spectral image dataset. The generated spectral images are validated as a data augmentation strategy by performing spectral image reconstruction based on compressive sensing and using RGB images.

Snapshot spectral imaging enables the acquisition of hyperspectral images (HSI) employing specialized optical systems, such as the coded aperture snapshot spectral imager (CASSI). Specifically, the CASSI system performs spatiospectral codification of light obtaining two-dimensional projected measurements, and these measurements are then processed by computational algorithms to obtain the desired spectral images. However, because HSIs often have a high spatial or spectral resolution, the sensing matrix related to the acquisition protocol becomes very large, leading to a high computational storage cost and long computation times. In this work, we propose an algebraic framework for computing the relevant operations in a tensorial form based on the nature of the codification protocol. We then test our framework against some comparison methods based on linear algebra decomposition, factorization, or block-operations, demonstrating that the proposed method is between 3 and 20 times faster than the best-competing method. Moreover, the gain becomes larger when the matrices become bigger, corresponding to realistic HSI sizes for spectral imaging applications. In extreme cases, our method can still operate when the competing methods stall due to memory shortage. (c) 2024 SPIE and IS&T

Learning-based reconstruction methods have shown state-of-the-art (SOTA) performance for compressive spectral imaging (CSI). However, the training setup dependence causes slight variations in the scene's statistical distribution or calibration errors in the sensing matrix producing poor testing reconstruction quality. Therefore, we propose a computational calibration methodology to improve any CSI reconstruction deep neural network (DNN) testing quality and robustness. Specifically, similar to transfer learning, we rethink the deep image prior framework to retrain a SOTA DNN for a particular CSI measurement without retraining the DNN from scratch. Experimental results show an average peak signal-to-noise ratio improvement of 5.5dB when slight variations in the CSI measurement and calibration errors in the sensing matrix are considered. Additionally, even when no variations are considered, the proposed calibration methodology improves in up to 2.3dB the reconstruction quality.

Spectral imaging collects and processes information along spatial and spectral coordinates quantified in discrete voxels, which can be treated as a 3D spectral data cube. The spectral images (SIs) allow the identification of objects, crops, and materials in the scene through their spectral behavior. Since most spectral optical systems can only employ 1D or maximum 2D sensors, it is challenging to directly acquire 3D information from available commercial sensors. As an alternative, computational spectral imaging (CSI) has emerged as a sensing tool where 3D data can be obtained using 2D encoded projections. Then, a computational recovery process must be employed to retrieve the SI. CSI enables the development of snapshot optical systems that reduce acquisition time and provide low computational storage costs compared with conventional scanning systems. Recent advances in deep learning (DL) have allowed the design of data-driven CSI to improve the SI reconstruction or, even more, perform high-level tasks such as classification, unmixing, or anomaly detection directly from 2D encoded projections. This work summarizes the advances in CSI, starting with SI and its relevance and continuing with the most relevant compressive spectral optical systems. Then, CSI with DL will be introduced, as well as the recent advances in combining the physical optical design with computational DL algorithms to solve high-level tasks.

Deep learning methods are state-of-the-art for spectral image (SI) computational tasks. However, these methods are constrained in their performance since available datasets are limited due to the highly expensive and long acquisition time. Usually, data augmentation techniques are employed to mitigate the lack of data. Surpassing classical augmentation methods, such as geometric transformations, GANs enable diverse augmentation by learning and sampling from the data distribution. Nevertheless, GAN-based SI generation is challenging since the high-dimensionality nature of this kind of data hinders the convergence of the GAN training yielding to suboptimal generation. To surmount this limitation, we propose low-dimensional GAN (LD-GAN), where we train the GAN employing a low-dimensional representation of the dataset with the latent space of a pretrained autoencoder network. Thus, we generate new low-dimensional samples which are then mapped to the SI dimension with the pretrained decoder network. Besides, we propose a statistical regularization to control the low-dimensional representation variance for the autoencoder training and to achieve high diversity of samples generated with the GAN. We validate our method LD-GAN as data augmentation strategy for compressive spectral imaging, SI super-resolution, and RBG to spectral tasks with improvements varying from 0.5 to 1 [dB] in each task respectively. We perform comparisons against the nondata augmentation training, traditional DA, and with the same GAN adjusted and trained to generate the full-sized SIs. The code of this paper can be found in https://github.com/romanjacome99/LD_GAN.git

Snapshot Spectral Imaging enables the acquisition of hyperspectral images (HSI) from 2D projected measurements employing specialized optical systems, such as the Coded Aperture Snapshot Spectral Imager (CASSI). Specifically, the CASSI system performs spatio-spectral codification of light obtaining 2D projected measurements. These measurements are then processed by algorithms to obtain the desired spectral images. Most traditional algorithms must compute an inverse matrix through decomposition, factorization, or block-operating a matrix related to the sensing protocol. However, since HSIs often have a high spatial or spectral resolution, the computation of an inverse matrix has a high computational cost. In this work, we propose an algebraic framework for computing the inverse matrix based on the nature of the codification protocol, accelerating its computation, in a tensorial form. Performed experiments from our proposed framework against some comparison methods based on linear algebra decomposition, factorization or block operations, show that the proposed framework is between 3 to 15 times faster than the best competing method, where the latter factor occurs when the matrices become bigger, which usually corresponds to realistic HSI sizes for spectral imaging applications.

The abundant spatial and angular information from light fields has allowed the development of multiple disparity estimation approaches. However, the acquisition of light fields requires high storage and processing cost, limiting the use of this technology in practical applications. To overcome these drawbacks, the compressive sensing (CS) theory has allowed the development of optical architectures to acquire a single coded light field measurement. This measurement is decoded using an optimization algorithm or deep neural network that requires high computational costs. The traditional approach for disparity estimation from compressed light fields requires first recovering the entire light field and then a post-processing step, thus requiring long times. In contrast, this work proposes a fast disparity estimation from a single compressed measurement by omitting the recovery step required in traditional approaches. Specifically, we propose to jointly optimize an optical architecture for acquiring a single coded light field snapshot and a convolutional neural network (CNN) for estimating the disparity maps. Experimentally, the proposed method estimates disparity maps comparable with those obtained from light fields reconstructed using deep learning approaches. Furthermore, the proposed method is 20 times faster in training and inference than the best method that estimates the disparity from reconstructed light fields.

Spectral image reconstruction from RGB images has emerged as a hot topic in the computer vision community due to easy-access and low-cost acquisition of the latter. The goal is to learn a non-linear mapping from 3-RGB bands to L spectral bands. With the growth of the available spectral datasets, this mapping has been learned using deep convolutional representations. However, these methods demand a large number of spectral images to train the net to obtain a good recovery. In contrast, the proposed process consists of a pre-training step where the weights of a convolutional neural network fit with a large amount of available RGB datasets without spectral mapping, taking into account the RGB system acquisition as a layer. Then, some layers of this pre-trained network are frozen to retrain it with the available spectral dataset to generate a spectral image with L bands. The proposed training scheme can be used with any pre-existing deep network that maps RGB to spectral images and it is here evaluated with a “U-net” architecture, and the RGB sensing is based on the Bayer filter pattern. The simulated and experimental data demonstrate the effectiveness of the proposed approach compared to training without transfer learning, showing a gain of up to 4 dB, with less spectral data.