Capturing images through transparent media often introduces superimposed reflections that substantially degrade image quality and distort polarization distributions, thereby hindering the performance of low-level vision tasks. By exploiting inherent differences in polarization states between transmitted and reflected light, polarization imaging provides a robust physical cue for resolving this ill-posed problem. However, existing polarization-based reflection-removal methods insufficiently exploit polarization cues and fail to effectively restore the polarization information of the transmission layer. To harness this potential, we propose a polarization information-guided cross dual-stream network (PICDS-Net) that effectively integrates polarization priors into a dual-branch architecture for simultaneous reflection removal and polarization recovery. In particular, the network established progressive, bidirectional, and polarization-guided feature interactions between reflection and transmission streams across multiple decoder stages, enabling effective removal through hierarchical refinement. To further enhance the fusion of polarization-sensitive features, we introduce a cross-branch attention block that adaptively fuses complementary features across streams by modeling their exclusivity and complementarity, ensuring that each stream focuses on its respective layer content while avoiding feature contamination. We also introduce a polarimetric consistency loss to ensure the plausibility of polarization, thereby improving the model's robustness and generalization. Experiments on real-world polarized image datasets demonstrated that PICDS-Net achieved state-of-the-art performance, producing cleaner transmission images with sharp details and recovering the scene's underlying polarization characteristics. By restoring physically consistent polarization characteristics of the transmission layer, this work highlights the advantages of accurate polarization recovery for real-world reflective imaging scenarios. Application-oriented demonstrations further show that reliable polarization restoration improves material perception and scene interpretation, offering valuable insights for advancing polarization-guided vision systems in complex environments.

Underwater imaging is essential for marine exploration, but wavelength-dependent absorption and scattering often degrade image quality, leading to color distortion and low contrast. Although recent diffusion models have shown promise in enhancing underwater images, most rely on iterative multi-step sampling and incur high computational costs, making them impractical for real applications such as ROV or AUV imaging. To address this, we present OSUIE (One-Step Underwater Image Enhancement), a novel framework built upon the latent diffusion model. We employ variational score distillation to condense the capabilities of a multi-step model into a singlestep network, enabling rapid inference without sacrificing visual quality. Furthermore, to mitigate frequency-domain artifacts in color information, we utilize a Fourier amplitude consistency constraint that ensures color fidelity. Extensive experiments on benchmarks and real-world videos demonstrate that OSUIE is at least 3.82 & times; faster than state-of-the-art diffusion models while delivering better visual performance. Ablation studies validate our design and additional qualitative tests on other settings indicate possible transfer potential beyond the primary task. Code and datasets are available at https://github.com/hyz-35/OSUIE.

Underwater imaging is fundamentally challenged by scattering and absorption, which severely degrade visibility in applications ranging from ocean observation and environmental monitoring to maritime security. Polarization-based descattering has shown strong potential, yet most existing approaches rely on traditional Stokes formulations and manual or heuristic parameter tuning, limiting robustness under complex conditions. To address this, we develop a polarization-guided Stokes descattering (PGSD) framework that integrates physical modeling with optimization. Specifically, three interpretable parameters-a scene-induced mixing weight, a polarized visibility factor, and a background scale-are introduced to compensate for reflection asymmetry, effective polarization observability, and global airlight scaling. A degree of linear polarization (DoLP)-gated airlight estimation is further designed to suppress scattering leakage, and a two-stage solver ensures stable parameter convergence. Experiments across multiple turbidity levels, in real seawater, and with non-polarized illumination scenarios demonstrate that PGSD yields sharper textures, higher contrast, and improved target discrimination compared with classical Stokes-based and intensity-only methods. These results highlight the effectiveness and robustness of the proposed framework, which could facilitate the practical deployment of polarimetric descattering in real-world applications.

Light propagation in aquatic environments is significantly impeded by scattering and absorption induced by suspended particulate matter, leading to severe image degradation and reduced visibility, especially under high-turbidity conditions. To address these challenges, numerous image enhancement and restoration algorithms have been developed. However, most existing methods are tailored to a specific water condition and lack adaptability to the diverse and dynamic optical properties of natural aquatic environments. In this paper, we propose an adaptive image enhancement method based on turbidity grading that effectively handles varying levels of water turbidity. First, a region-weighted clarity evaluation is introduced to accurately assess image sharpness and contrast under uneven lighting and scattering conditions. Based on this clarity measure, we establish a mapping between visual clarity and turbidity levels, categorizing the water environment into three grades: low, moderate, and high turbidity. Subsequently, different enhancement strategies are adaptively applied to images according to their estimated turbidity category. Extensive experiments conducted on both public underwater datasets and self-collected scenes demonstrate that the proposed method achieves superior visual quality and quantitative performance compared with several state-of-the-art algorithms. The results confirm that our approach can effectively remove haze, restore color fidelity, enhance contrast and texture details, and preserve overall naturalness across diverse underwater conditions.

Exploring the ocean’s vast, water-related environment, covering over 70

Infrared small target detection (IRSTD) remains a challenging task due to the low signal-to-noise ratio and significant background interference, which make it difficult to preserve fine target details. Conventional neural networks, constrained by their local receptive fields, often struggle to separate complete target details from high-frequency noise across channels, resulting in suboptimal performance. To address this limitation, we propose a parametric manifold network, i.e., PMNet, that integrates frequency-domain analysis with manifold learning. Based on the wavelet decomposition, PMNet separates high-frequency features from the spatial domain to encode high-frequency target details across channels; accordingly, the manifold-aware module, built upon parametric uniform manifold approximation and projection, captures non-local cross-channel high-dimensional features. This design effectively preserves the geometric structure of the target while suppressing noise, thus achieving more accurate edge and contour segmentation. Additionally, a low-frequency feature block is introduced to refine structural features, enhancing discrimination between targets and backgrounds. Extensive experiments on three IRSTD datasets demonstrate that the PMNet outperforms state-of-the-art methods in both accuracy and robustness, highlighting the effectiveness of integrating wavelet decomposition with manifold learning in IRSTD. Open-source implementations will be available at https://github.com/Lianruoqi/PMNet.

Underwater polarization imaging is a promising method as applied to sensing the ocean. However, existing methods primarily focus on restoring intensity images while neglecting the retrieval of polarization information itself. As an additional information dimension, polarization can significantly enhance imaging performance in various underwater scenarios, such as target detection and autonomous navigation. This paper integrates polarization information into a modified dual-stream diffusion model for the joint restoration of intensity and polarization information in turbid underwater conditions. To bridge the dual diffusion processes, a descattering network with polarization restoration sub-networks is designed, which intuitively takes polarization information as input and fully exploits informative polarization features. Polarization correlation is adopted as a useful prior to recalibrate the polarization features of the sub-networks, which is achieved by polarization hybrid attention blocks. Additionally, a specially designed loss function that considers polarization criteria enables the diffusive image restoration process to focus on descattering-related polarization features and improve the accuracy of the restored polarization information. Comprehensive experiments demonstrate that the proposed method outperforms state-of-the-art underwater imaging methods in both polarization image quality and polarization information restoration accuracy. Furthermore, sea trials validate its robust imaging performance under real marine conditions. Combined with several practical application cases, this work highlights the advantages of the restored polarization information for addressing real-world underwater tasks.

Infrared small target detection (IRSTD) remains challenging due to the scarcity of useful target cues and the presence of severe background clutter. Most current methods rely on conventional feature learning and local interaction modeling, where features are represented in Euclidean space. However, such designs may still be limited in describing the subtle differences of weak targets and the contextual relations between targets and backgrounds. To address these limitations, we propose LoHGNet, an IRSTD network that integrates Lorentz geometric encoding with high-order relation learning. By introducing Lorentz manifold based feature learning, LoHGNet offers a different feature representation from conventional IRSTD methods and provides new discriminative cues for IRSTD. Specifically, a Lorentz encoding branch is constructed with the Geometric Attention Guided Lorentz Residual Convolution Module (GA-LRCM) to perform feature modeling under hyperbolic geometric constraints and enhance the hierarchical geometric representation capability of weak targets. Subsequently, the hyperbolic features are mapped into the Euclidean tangent space through logarithmic mapping, and a High-Order Relation Learning Module (HORL) is designed to model the high-order contextual dependencies between targets and backgrounds via hypergraph construction, thereby improving target discrimination in complex backgrounds. Experimental results on three datasets demonstrate that the proposed LoHGNet achieves competitive performance in both detection accuracy and adaptability to complex scenes. The code will be available at https://github.com/Kingwin97.

Due to rapid advances in deep learning, many polarimetric image denoising networks have been developed and achieved promising results. However, these methods are based on general network architectures that do not fully exploit problem-specific knowledge, leading to over-smoothing results and poor generalization. Inspired by the non-local, which is an effective prior for image restoration, we propose a cube matching convolutional neural network to incorporate non-local operations into denoising models for polarimetric images. Specifically, the cube matching technique allows the denoising network to simultaneously exploit the non-local correlation and polarization relationship between the corresponding voxels of similar cubes. Rather than applying self-similarity directly in an isolated manner, the proposed cube matching module can be flexibly integrated into existing deep networks by combining with 3D convolution, achieving an effect equivalent to non-local means. This design enhances the generalization ability against various types and levels of noise. Experimental results show that using cube matching operations significantly improves denoising performance.

To achieve efficient denoising of polarization images without losing their original polarization characteristics,this paper proposes a polarization image denoising algorithm based on masked pre-training.The algorithm first performs pre-training on a clear polarization image dataset by applying masks,using the parameters obtained from the pre-trained model as initial parameters for self-supervised masked training.Noisy polarization images are then input for iterative training.During each iteration,predictions for the masked pixels are extracted,and an exponential moving average strategy is employed to generate the final prediction of the clean image.Experimental results demonstrate that the proposed algorithm not only achieves direct denoising of polarization images but also delivers superior visual performance in restoring polarization information.Additionally,its robustness is validated across different noise patterns and noise levels.

Underwater laser ranging technology is a crucial approach for ocean exploration and underwater sensing,with broad applications in marine engineering,environmental monitoring,and resource development.Owing to its advantages of rapid and accurate measurement,this technology demonstrates highly efficient detection capabilities in complex aquatic environments,providing strong support for acquiring high-resolution underwater information.This study first introduces the principles and recent technological advancements of four commonly used underwater laser ranging methods,including time-of-flight,phase detection,triangulation,and photon counting.It then reviews the typical application progress of this technology,followed by a comparative analysis of the advantages and limitations of each method in underwater detection.Finally,future development trends and potential applications of underwater laser ranging are discussed.

In this paper we present a magnetic probe structure based on magneto-fluidic and laser internal cavity modulation. We used a single-mode-no-core-single-mode fiber structure coated by a magnetic fluid as a sensing element, and inserted the sensing element in the inner cavity of a laser. Highly sensitive, high signal-to-noise ratio (SNR), narrow half-height width (FWHM) sensing signals have been obtained using intra-laser cavity modulation. The magnetic detection sensitivity of the sensor was improved, and a sensing system with a magnetic detection sensitivity of 1.76 x 10(-3) mw/Oe and a linearity of 99.709% was achieved. The temperature sensitivity is 0.035 mu W/degrees C. The limit error of the magnetic field sensor is +/- 0.006035%. Meanwhile, output signal index of the sensing system is FWHM<40pm. (c) 2025 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement

Full-polarization light detection and ranging (LiDAR) is a powerful remote sensing technology that provides comprehensive information about object distance, material properties, surface texture, and orientation. However, existing full-polarization LiDAR systems typically rely on time-division or amplitude-division polarization schemes, requiring multiple measurements and computational processing to reconstruct the full-Stokes vector. This limits real-time performance and increases system complexity. Here, we present a compact full-polarization LiDAR, termed VRP-LiDAR, which leverages the spatial polarization modulation capability of a vortex retarder (VR) in conjunction with a polarization camera and a photodetector. This approach enables real-time, single-shot acquisition of both distance information and full-Stokes polarization data without requiring multiple exposures or moving parts. Experimental validation confirms the effectiveness of VRP-LiDAR, demonstrating its potential for compact and robust polarization-based target detection in remote sensing applications.

Optical imaging and vision technology have become crucial research topics in the field of underwater and ocean scenes. These technologies play a vital role in advancing underwater exploration, scientific research, and smart aquaculture. Despite their importance, optical systems face substantial challenges in complex water environments, especially in turbid conditions where light absorption and scattering by water and suspended particles drastically reduce the quality of captured images. This degradation highlights the need for sophisticated image restoration algorithms that can effectively recover the quality of images impaired by turbid water conditions. Cascaded frameworks, which incorporate multiple modules, including that of color correction, detail enhancement, and information fusion, have emerged as particularly efficacious, offering significant improvements over traditional direct or end-to-end restoration methods. This review elaborates on the comparative benefits of various cascading structures—serial, parallel, and hybrid—and examines different restoration algorithms that utilize intrinsic image features, physical models, prior knowledge, and deep learning technologies. Furthermore, it provides a comprehensive overview of publicly available datasets and evaluation metrics, both referential and non-referential, which are crucial for developing, validating, and optimizing underwater image restoration algorithms. Additionally, by conducting experimental comparative studies of several open restoration algorithms, the effectiveness and applicability of different methods are validated. Finally, the review proposes potential research directions, aiming to deliver valuable insights to scholars in underwater application and computer vision, thereby inspiring innovative research initiatives that could further enhance the capabilities of underwater optical imaging.

Underwater object detection (UOD) is pivotal for applications in aquaculture, marine resource exploration, and environmental monitoring. However, relying solely on vision-guided image enhancement techniques as a preprocessing step for UOD is inadequate to address the prevalent degradation challenges in underwater imaging. To overcome the limitation, this paper proposes an unified adaptive enhancement and detection network (UAED-Net), which enhances the texture information of detection features through progressively integrating enriched features generated by an enhancement module; thereby improving the overall performance of the detector. Specifically, UAED-Net incorporates a 2nd-order Sobel operator within the detection-aware feature enhancement module. The operator's elevated central weights enhance its sensitivity to subtle texture variations and structurally complex edges and corners in images. By processing images across horizontal, vertical, and diagonal directions, it enables the extraction of comprehensive texture features. Joint training of the enhancement module and the UOD network provides auxiliary discriminative information, further strengthening the network's predictive capabilities. To achieve effective integration of enhanced and detection features, as well as cross-scale feature fusion across different dimensions, a mutual adaptive feature fusion model is introduced. This model enhances the spatial representation of object features, enabling the detection branch to learn richer target information and optimize detection performance. Experimental results on four challenging UOD datasets demonstrate that the proposed UAED-Net achieves superior performance, highlighting its effectiveness in addressing the complexities of underwater imaging. Link to open-source code: https://github.com/LeeBincheng/UAED-Net.

In the field of infrared small target detection (IRSTD), targets generally exhibit dim characteristics and are difficult to distinguish from background clutter. Learning-based methods enhance feature representation through layer-by-layer propagation, but the sparse target information often diminishes. To address this, we propose DWTFreqNet, a network that splits input data to enhance both local saliency and global contextual differences. It incorporates complementary feature extraction modules designed to match the data distribution characteristics. Specifically, it first utilizes the discrete wavelet transform (DWT) to decompose the input data into low- and high-frequency components. For the low-frequency part, which carries key target information, we apply component-differential dense connections and DWT-based downsampling to maintain feature integrity. For the high-frequency part, rich in target-background contrast, an adaptive wavelet guidance mechanism (AWGM) optimizes multicomponent fusion via adaptive weighting, while a layer-wide discrepancy relationship capture (LDRC) module enhances target discrimination by linking multiscale feature maps. Comparative experiments on public datasets demonstrate its superiority over state-of-the-art (SOTA) methods. The code will be available at https://github.com/Kingwin97/DWTFreqNet

Computational imaging technology,by integrating front-end optical design with back-end signal processing algorithms,has broken through the physical limitations of conventional imaging techniques and pioneered a new imaging paradigm.This technology demonstrates unique advantages in the imaging field due to its flexible system degrees of freedom,multi-dimensional information acquisition capability,and intelligent processing characteristics.At its core lies solving the inverse problem of imaging systems.However,as problem complexity increases,traditional physics-based methods face dual challenges of low computational efficiency and insufficient modeling accuracy. The introduction of deep learning offers new solutions to this challenge,yet purely data-driven approaches remain constrained by inherent limitations such as heavy reliance on annotated data and inadequate generalization capabilities.Model-driven methods enhance interpretability by incorporating physical principles and prior knowledge,but demonstrate limited modeling capacity for complex problems.Generative Artificial Intelligence(AI)combines the strengths of both approaches by learning data distributions to provide adaptive prior information.Current mainstream generative models include Flow-based Model,Variational Autoencoder,and Generative Adversarial Network,while the emerging Diffusion Model demonstrates unique advantages by accomplishing image generation,super-resolution,and other tasks without requiring complex constraints.The integration of generative AI with computational optical imaging presents synergistic opportunities:generative AI enables advanced image reconstruction algorithms that significantly enhance imaging quality,while computational imaging supplies high-fidelity multimodal training data that improves the authenticity and diversity of AI-generated outputs. The review of generative prior learning applications in single-pixel imaging,scattering imaging,lensless imaging,and single-exposure compressive imaging demonstrates that intelligent computational optical imaging systems fundamentally center on optoelectronic architectures that provide data sources through encoding/decoding mechanisms.The physical properties of light determine modeling paradigms,delivering multidimensional physical information for computational imaging.Sensor limitations induce signal dimensionality reduction,creating specialized data characteristics.Within this framework,image reconstruction emerges as the mathematical core of computational imaging,synthesizing physical models with data attributes.Modern generative AI techniques leverage large-scale prior learning to provide effective constraints and computational assistance throughout measurement recovery and image reconstruction processes. Future computational optical imaging will advance through three key breakthroughs.At the data level,the focus will be on developing universal large-scale models and standardized databases that integrate multimodal imaging data,establishing a robust foundation for the field.At the algorithmic level,researchers will create generative deep learning models incorporating the physical properties of optical components,embedding these physical constraints into algorithm design to enhance reconstruction accuracy.At the application level,closed-loop systems implementing"data generation-imaging reconstruction-demand guidance"will be established,achieving intelligent imaging through hardware-software co-optimization.These advances will propel computational optical imaging toward smarter and more efficient paradigms,driving comprehensive improvements from theory to practical implementation.

We implement a four-branch denoiser for polarimetric imaging LiADR that uses a fractal superconducting nanowire single-photon detector. This denoiser permits the polarimetric imaging LiDAR to acquire images with decent qualities in photon-starved conditions and relatively short acquisition time.

Infrared small target detection (IRSTD) remains challenging due to the weak spatial features of targets and their susceptibility to background clutter. Recent studies have improved detection performance through the embedding of additional spatial representations. However, these feature prompting methods rely on discretely sampled feature spaces, which weaken high-frequency information and consequently limit their representational efficiency. To overcome this, we propose NeRI, a network that leverages the potential of implicit neural representations (INRs) through a continuous formulation to learn mappings from spatial coordinates to the high-frequency structural representations of targets. Specifically, these mappings are realized through INR blocks (INRBs) integrated into different encoder layers, providing continuous spatial guidance from multiscale inputs and enabling more accurate localization and distinction. In addition, to better model the distinction between foreground and background, we construct a hybrid U-shaped block (HUB) that combines a U-shaped transformer block (UTB) and multiscale convolution block (MCB). The UTB component effectively increases network depth and facilitates long-range dependency modeling across different scales, while the MCB employs convolutions with varying receptive fields to capture fine-grained local information, thereby enabling the two components to fully exploit their complementary strengths. Finally, we propose a simple yet effective spatial-semantic fusion (SSF) module that reweights and integrates spatial information from diverse layers to enhance the expressive power of the features. The proposed NeRI offers a robust solution for the accurate separation of targets from backgrounds. Experimental validation, conducted on three public datasets (i.e., NUDT-SIRST, NUAA-SIRST, and IRSTD-1K), demonstrates the superior performance of NeRI compared to other methods. Open-source implementations will be available at https://github.com/Shangwei-Deng/NeRI

Unmanned Underwater Vehicles (UUVs) struggle to acquire high-accuracy images in complex underwater environments, where overreliance on physical priors hampers AI-based image processing. Color polarization imaging, though underutilized, captures multi-dimensional scene details, offering significant potential for intelligent marine robotics. We propose a lightweight, end-to-end invertible network, termed CPSENet, which leverages fractional calculus operators for de-scattering and fine-grained enhancement of turbid underwater polarization images. Specifically, we introduce a globally adaptive dynamic fractional convolution residual module. It exploits the memory and non-local properties inherent in fractional calculus operators and integrates the degree of linear polarization, which encapsulates the geometric characteristics of target scenes. This approach enhances long-range modeling and preserves global edge information during the image enhancement, thereby reducing the risk of losing high-frequency details. Additionally, CPE-INN, a dense connection invertible network fusion strategy, ensures lossless transmission of high-order features, minimizing color distortion and enhancing multi-modal feature integration. This approach minimizes color distortion while facilitating the efficient integration of multimodal features. Finally, we construct a comprehensive, large-scale, real-world underwater target color polarization dataset, termed PCE-Dataset. Experiments under active non-uniform illumination demonstrate superior image restoration and compact model size, offering an efficient solution for UUV imaging detection.