Spectral images are highly valuable for various applications, including environmental monitoring and precision agriculture. However, the high cost of specialized sensors limits the wide use of this technology in numerous applications. Current alternatives to acquire high spatial-spectral resolution spectral images, like Single-Pixel Imaging (SPI) enhanced with Deep Optical Coding Design (DOCD), have limitations due to their non-feedback optical designs, leading to limited image quality, with optimal performance achieved only for the specific scenes used during training. This work reformulates the DOCD framework to handle the scene-driven ordering of the Hadamard basis within the SPI architecture for spectral imaging. Taking into account that SPI usually acquires hundreds of snapshots, our approach introduces a scene-driven ordering of the Hadamard matrix for flexible SPI modulation pattern selection based on scene characteristics in an end-to-end optimization. Simulations on spectral datasets and real test-bed acquisitions demonstrate the effectiveness of the proposed method in improving the quality of VIS and NIR spectral images compared to fixed designs.

Modulo-Imaging (MI) offers a promising alternative for expanding the dynamic range of images by resetting the signal intensity when it reaches the saturation level. Subsequently, high-dynamic range (HDR) modulo imaging requires a recovery process to obtain the HDR image. MI is a non-convex and ill-posed problem where recent recovery networks suffer in high-noise scenarios. In this work, we formulate the HDR reconstruction task as an optimization problem that incorporates a deep prior and subsequently unrolls it into an optimization-inspired deep neural network. The network employs a lightweight convolutional denoiser for fast inference with minimal computational overhead, effectively recovering intensity values while mitigating noise. Moreover, we introduce the Scaling Equivariance term that facilitates self-supervised fine-tuning, thereby enabling the model to adapt to new modulo images that fall outside the original training distribution. Extensive evaluations demonstrate the superiority of our method compared to state-of-the-art recovery algorithms in terms of performance and quality.

Self-supervised image denoising methods have traditionally relied on either architectural constraints or specialized loss functions that require prior knowledge of the noise distribution to avoid the trivial identity mapping. Among these, approaches such as Noisier2Noise or Recorrupted2Recorrupted, create training pairs by adding synthetic noise to the noisy images. While effective, these recorruption-based approaches require precise knowledge of the noise distribution, which is often not available. We present Learning to Recorrupt (L2R), a noise distribution-agnostic denoising technique that eliminates the need for knowledge of the noise distribution. Our method introduces a learnable monotonic neural network that learns the recorruption process through a min-max saddle-point objective. The proposed method achieves state-of-the-art performance across unconventional and heavy-tailed noise distributions, such as log-gamma, Laplace, and spatially correlated noise, as well as signal-dependent noise models such as Poisson-Gaussian noise.

Object detection precision is crucial for ensuring the safety and efficacy of autonomous driving systems. The quality of acquired images directly influences the ability of autonomous driving systems to correctly recognize and respond to other vehicles, pedestrians, and obstacles in real-time. However, real environments present extreme variations in lighting, causing saturation problems and resulting in the loss of crucial details for detection. Traditionally, High Dynamic Range (HDR) images have been preferred for their ability to capture a broad spectrum of light intensities, but the need for multiple captures to construct HDR images is inefficient for real-time applications in autonomous vehicles. To address these issues, this work introduces the use of modulo sensors for robust object detection. The modulo sensor allows pixels to `reset/wrap' upon reaching saturation level by acquiring an irradiance encoding image which can then be recovered using unwrapping algorithms. The applied reconstruction techniques enable HDR recovery of color intensity and image details, ensuring better visual quality even under extreme lighting conditions at the cost of extra time. Experiments with the YOLOv10 model demonstrate that images processed using modulo images achieve performance comparable to HDR images and significantly surpass saturated images in terms of object detection accuracy. Moreover, the proposed modulo imaging step combined with HDR image reconstruction is shorter than the time required for conventional HDR image acquisition.

Drone detection under high-illumination conditions remains a critical challenge due to sensor saturation, which degrades visual information and limits the performance of conventional detection models. A promising alternative to overcome this issue is modulo imaging, an approach based on modulo-ADCs that reset pixel intensities upon reaching a predefined saturation threshold, thus avoiding saturation loss. This work presents a methodology based on fine-tuning a detection model using modulo images, allowing accurate object detection without requiring High Dynamic Range (HDR) image reconstruction. Additionally, an optional reconstruction stage using the Autoregressive High-order Finite Difference (AHFD) algorithm is evaluated to recover high-fidelity HDR content. Experimental results show that the fine-tuned model achieves F1-scores above 96% across different illumination levels, outperforming saturated and raw modulo inputs, and approaching the performance of ideal HDR images. These findings demonstrate that fine-tuning with modulo data enables robust drone detection while reducing inference time, making the reconstruction process optional rather than essential.

Modulo imaging enables high dynamic range (HDR) acquisition by cyclically wrapping saturated intensities, but accurate reconstruction remains challenging due to ambiguities between natural image edges and artificial wrap discontinuities. This work proposes a learning-based HDR restoration framework that incorporates two key strategies: (i) a scale-equivariant regularization that enforces consistency under exposure variations, and (ii) a feature lifting input design combining the raw modulo image, wrapped finite differences, and a closed-form initialization. Together, these components enhance the network's ability to distinguish true structure from wrapping artifacts, yielding state-of-the-art performance across perceptual and linear HDR quality metrics.

Cocoa beans fermentation is a key process that defines the sensory quality of cocoa-based products such as chocolate. However, current methods for estimating the Fermentation Index (FI), such as cut tests and chemical analysis are destructive, delayed, and costly. Recent approaches have explored nondestructive techniques based on spectral information and specific band selection, but these methods often prioritize statistical performance while ignoring physical constraints, selecting unrealistic spectral bands. To address this, we propose a physically feasible and efficient framework for FI estimation that combines spectral dimensionality reduction with a fermentation prediction model, aligned with commercial optical hardware limitations. The method consists of designing a set of Gaussian Optical Bandpass Filters, optimized in terms of central wavelength and bandwidth, and integrating them into a learning pipeline jointly trained with a lightweight neural network regression model. The proposed model achieves an R-2 of 0.959 using six filters, closely matches the full spectrum baseline (R-2 = 0.963) while reducing the input dimensionality by over 98%, outperforming the latest band selection methods in R-2 and MAE scores.

High dynamic range (HDR) imaging is vital for capturing the full range of light tones in scenes, essential for computer vision tasks such as autonomous driving. Standard commercial imaging systems face limitations in capacity for well depth, and quantization precision, hindering their HDR capabilities. Modulo imaging, based on unlimited sampling (US) theory, addresses these limitations by using a modulo analog-to-digital approach that resets signals upon saturation, enabling estimation of pixel resets through neighboring pixel intensities. Despite the effectiveness of (US) algorithms in one-dimensional signals, their optimization problem for two-dimensional signals remains unclear. This work formulates the US framework as an autoregressive ℓ_2 phase unwrapping problem, providing computationally efficient solutions in the discrete cosine domain jointly with a stride removal algorithm also based on spatial differences. By leveraging higher-order finite differences for two-dimensional images, our approach enhances HDR image reconstruction from modulo images, demonstrating its efficacy in improving object detection in autonomous driving scenes without retraining.

Image acquisition in low-light environments is fundamentally challenging due to the photon-limited nature of the scene, which results in severe noise and incomplete color information. Imaging sensors operating under such conditions require robust post-processing to recover visually coherent, full-color images. In these conditions, the photon arrival process can be modeled as a Poisson distribution, which introduces noise that complicates image reconstruction. Furthermore, the use of a color filter array leads to missing color information at each pixel, which exacerbates the challenge. As a result, denoising and demosaicing become ill-posed and interdependent tasks. We propose a self-supervised method that jointly addresses denoising and demosaicing under low-light conditions without requiring clean reference images. Our approach achieves a PSNR higher by 2.0 dB compared to best state-of-the-art methods at gain of 20 and is close to the supervised method.

Recorrupted-to-Recorrupted (R2R) has emerged as a methodology for training deep networks for image restoration in a self-supervised manner from noisy measurement data alone, demonstrating equivalence in expectation to the supervised squared loss in the case of Gaussian noise. However, its effectiveness with non-Gaussian noise remains unexplored. In this paper, we propose Generalized R2R (GR2R), extending the R2R framework to handle a broader class of noise distribution as additive noise like log-Rayleigh and address the natural exponential family including Poisson, Gamma and Binomial noise distributions, which play a key role in many applications including lowphoton imaging and synthetic aperture radar. We show that the GR2R loss is an unbiased estimator of the supervised loss and that the popular Stein's unbiased risk estimator can be seen as a special case. A series of experiments with Gaussian, Poisson, and Gamma noise validate GR2R's performance, showing its effectiveness compared to other selfsupervised methods.

DeepInverse is an open-source PyTorch-based library for solving imaging inverse problems. The library covers all crucial steps in image reconstruction from the efficient implementation of forward operators (e.g., optics, MRI, tomography), to the definition and resolution of variational problems and the design and training of advanced neural network architectures. In this paper, we describe the main functionality of the library and discuss the main design choices.

This paper introduces a plug-and-play algorithm for enhancing compressive spectral imaging (CSI) through the integration of both a quadratic envelope (QE) regularizer and a deep prior. Our method employs the QE-based regularizer to foster a low-rank structure in conjunction with deep priors, synergistically integrated within a Plug-and-Play (PnP) framework. The distinct advantage of our chosen QE-regularizer is its propensity for uncovering low-rank solutions devoid of bias, distinguishing it from the nuclear norm. Through this fusion of QE and deep priors, we harness the complementary strengths of both techniques, resulting in a mutually reinforcing effect for CSI.

The Hadamard Single-Pixel Imaging (HSPI) has emerged as a cost-effective approach for spectral image acquisition, capturing inner products between scenes and coded apertures chosen as rows of the Hadamard matrix. Current strategies employ ordering methodologies to reduce acquisition time by selecting a subset of Hadamard rows. However, they encounter limitations in fixed sensing, particularly in preserving image quality, as each spectral image may have a unique subset of coded apertures that retains the most spatial information. This work tackles these challenges by introducing an adaptive acquisition protocol for HSPI. Our two-step approach seamlessly integrates fixed and adaptive sensing stages, optimizing coded aperture selection through a binary classification model in the 2D Hadamard spectrum. The comparative analysis demonstrates the superiority of our adaptive protocol over existing ordering methodologies, offering a flexible solution to overcome limitations associated with fixed acquisitions.

Commercial sensors often suffer from overexposure in bright regions, leading to signal clipping and information loss because of saturation. Existing solutions involve either employing logarithmic irradiance response sensors or capturing multiple shots from different saturation levels. However, these approaches can be complex or rely on static scenes, limiting their effectiveness in fully addressing the saturation problem. A promising solution is the use of unsaturated sensors, also known as modulo cameras, which employ an array of self-reset pixels to wrap the signal when it reaches the saturation level. The resulting image exhibits a noisy and discontinuous shape, requiring an unwrapping algorithm to obtain a smooth and continuous representation of the scene. We propose a deep plug-and-play algorithm that combines model-based optimization with a deep denoiser. By leveraging the spatial correlation of the scene within the close solution of an unwrapping step, our approach successfully unwraps the continuous values while simultaneously reducing noise. Extensive evaluations show the superiority of our method compared to state-of-the-art unwrapping and unmodulo algorithms in terms of reconstruction quality.

Binary Neural Networks emerged as a cost-effective and energy-efficient solution for computer vision tasks by binarizing either network weights or activations. However, common binary activations, such as the Sign activation function, abruptly binarize the values with a single threshold, losing fine-grained details in the feature outputs. This work proposes an activation that applies multiple thresholds following dithering principles, shifting the Sign activation function for each pixel according to a spatially periodic threshold kernel. Unlike literature methods, the shifting is defined jointly for a set of adjacent pixels, taking advantage of spatial correlations. Experiments over the classification task using both grayscale and RGB datasets demonstrate the effectiveness of the designed dithering Sign activation function as an alternative activation for binary neural networks, without increasing the computational cost. Further, DeSign balances the preservation of details with the efficiency of binary operations.

In this paper, we introduce an efficient algorithm for generating specific Hadamard rows, addressing the memory demands of pre-computing the entire matrix. Leveraging Sylvester's recursive construction, our method generates the required i-th row on demand, significantly reducing computational resources. The algorithm uses the Kronecker product to construct the desired row from the binary representation of the index, without creating the full matrix. This approach is particularly useful for single-pixel imaging systems that need only one row at a time.

Hadamard single-pixel imaging (HSI) is a promising sensing approach for acquiring spectral images in the near-infrared spectrum with high spatial resolution and fast recovery times due to the efficient invertible properties of the Hadamard matrix. The potential of the HSI system is diminished because of the large number of required measurements which implies long acquisition times. Recent advances proposed optimizing the HSI sensing matrix structure based on a superpixels map estimated from a side-information acquisition of the scene, reducing the number of required measurements. However, these matrix designs are detached from the recovery task, which falls on a sub-optimal strategy. In this work, we proposed an adaptive end-to-end sensing methodology for the HSI sensing matrix design based on deep superpixels estimation by coupling the sensing and recovery of the near-infrared spectral images. Experimental results show the superiority of the proposed sensing methodology compared with state-of-art sensing design schemes.

Chronic wounds are a latent health problem worldwide, due to high incidence of diseases such as diabetes and Hansen. Typically, wound evolution is tracked by medical staff through visual inspection, which becomes problematic for patients in rural areas with poor transportation and medical infrastructure. Alternatively, the design of software platforms for medical imaging applications has been increasingly prioritized. This work presents a framework for chronic wound tracking based on deep learning, which works on RGB images captured with smartphones, avoiding bulky and complicated acquisition setups. The framework integrates mainstream algorithms for medical image processing, including wound detection, segmentation, as well as quantitative analysis of area and perimeter. Additionally, a new chronic wounds dataset from leprosy patients is provided to the scientific community. Conducted experiments demonstrate the validity and accuracy of the proposed framework, with up to 84.5% in precision.

The Deep Image Prior (DIP) technique has been successfully employed in Compressive Spectral Imaging (CSI) as a non-data-driven deep model approach. DIP methodology updates the deep network’s weights by minimizing a loss function that considers the difference between the measurements and the forward operator of the network’s output. However, this method often yields local minima as all the measurements are evaluated at each iteration. This paper proposes a stochastic deep image prior (SDIP) approach, which stochastically trains DIP networks using random subsets of measurements from different CSI sensors in a CSI fusion (CSIF) setting, resulting in the improvement of the convergence through stochastic gradient descent optimization. The proposed SDIP method improves upon the deterministic DIP and requires less computational time since fewer forward operators are required per iteration. The SPID method provides comparable performance against the state-of-the-art CSF techniques based on supervised data-driven and unsupervised methods, achieving up to 5 dB in the reconstruction.

Deep learning models are state-of-the-art in compressive spectral imaging (CSI) recovery. These methods use a deep neural network (DNN) as an image generator to learn non-linear mapping from compressed measurements to the spectral image. For instance, the deep spectral prior approach uses a convolutional autoencoder (CAE) network in the optimization algorithm to recover the spectral image by using a non-linear representation. However, the CAE training is detached from the recovery problem, which does not guarantee optimal representation of the spectral images for the CSI problem. This work proposes a joint non-linear representation and recovery network (JR2net), linking the representation and recovery task into a single optimization problem. JR2net consists of an optimization-inspired network following an alternating direction method of multipliers (ADMM) formulation that learns a non-linear low-dimensional representation and simultaneously performs the spectral image recovery, trained via the end-to-end approach. Experimental results show the superiority of the proposed method with improvements up to 2.57 dB in peak signal-to-noise ratio (PSNR) and performance around 2000 times faster than state-of-the-art methods.