
To address the degradation of intensity-based defect representations caused by scattering, low illumination, and local overexposure in complex underwater environments, as well as the spatially varying reliability of polarization information, we propose a Polarization Reliability-guided Multimodal Fusion Network, termed PRMF-Net, for underwater bridge-pier defect segmentation. PRMF-Net first employs a polarization reliability prior generation module to derive polarization structural features, a boundary prior, and a spatial reliability map from four-angle polarization observations, thereby enhancing informative polarization responses while suppressing unstable ones. An asymmetric modality-decoupled encoder then preserves modality-specific characteristics in shallow layers and aligns high-level semantics across intensity and polarization modalities. Adaptive hierarchical fusion and progressive decoding are further introduced to jointly model intensity textures, reliable polarization features, and fine-grained defect boundaries. A multi-source polarimetric defect dataset is constructed from self-collected indoor data, improved CUT-synthesized data, and real underwater bridge-pier data, covering cracks, spalling, exposed tendons, and holes. PRMF-Net achieves mIoU scores of 82.6%, 80.4%, and 74.6% on the three datasets, outperforming the best single-modality methods by 5.8, 4.9, and 8.4 percentage points and the best multimodal methods by 2.8, 1.7, and 3.9 percentage points, respectively. Ablation studies confirm the effectiveness of polarization reliability modeling, modality-decoupled encoding, and hierarchical fusion. The proposed method improves defect-region continuity and boundary localization under complex underwater conditions.
With the advancement of hardware and algorithms, Gray-code-based phase-shifting profilometry have undergone huge advances over the past two decades. However, phase unwrapping errors and low projection efficiency have not been resolved, especially in dynamic scenarios. In this work, we proposed an algorithm to generate Gray code from periodic wave to avoid extra picture projection. Then, based on the algorithm, we proposed a Generative Complementary Gray code method (GCG) to avoid jump errors. In the GCG method, a periodic waveform is constructed by fully exploiting the projected images, and the (n+1)th Gray-code pattern can be generated without compromising noise robustness or reconstruction accuracy. Moreover, owing to its high error tolerance, the method can apply time-overlapping strategy to achieve efficient 3D reconstruction in dynamic scenes. Finally, in order to verify the performance of the method, we setup a 3D reconstruction system with 500(2000/4) Hz to capture the process of a rotating fan blade and the collapse of building block. And a series of comparative experiments are conducted in complex scenarios. The results show that the proposed method exhibits good robustness in dynamic scenes.
Metasurfaces enable precise manipulation of the amplitude, phase, and polarization of light fields at the subwavelength scale, serving as a core platform for next-generation planar optics. Electron-beam lithography excels at high-resolution prototype fabrication, while deep-ultraviolet lithography and nanoimprint lithography dominate wafer-level planar manufacturing. However, these conventional routes still face limitations in constructing true three-dimensional (3D) structures, integrating onto non-planar substrates, and achieving mask-free rapid patterning. Benefiting from nonlinear absorption and localized energy deposition, ultrafast laser micro/nano-fabrication provides a flexible and complementary technological alternative. This review systematically summarizes recent advances in three primary categories of ultrafast-laser-fabricated metasurfaces: laser-induced periodic surface structures (LIPSS), ultrafast laser direct writing (ULDW, encompassing direct ablation, two-photon polymerization, and in-volume modification), and spatial-light-modulator (SLM)-based beam-shaping printing. The multi-timescale laser–matter interaction mechanisms, complete fabrication process chains, and technical bottlenecks corresponding to each route are analyzed. Organized by device functionality, these fabrication routes are mapped onto typical applications including metalenses, holographic imaging and light-field encoding devices, beam deflectors, structural colors, and mid-infrared perfect absorbers. Quantitative comparisons are conducted across simulated and experimentally measured performance, resolution, 3D structural freedom, processing throughput, and material compatibility. Our analysis indicates that although these ultrafast laser routes exhibit distinct complementarity, most current laser-fabricated metasurface devices remain at the small-area proof-of-principle stage, lagging behind mature lithography processes in wafer-level uniformity, batch-to-batch consistency, and per-unit cost. Finally, future research directions are outlined across four dimensions: improving processing throughput and consistency, expanding high-refractive-index and phase-change functional materials, integrating hybrid multi-process routes, and systematically verifying device-level reliability—underscoring that ultrafast laser fabrication is not intended to replace mature nanomanufacturing technologies, but rather to provide a competitive complementary route in niche scenarios such as complex 3D topographies, conformal integration on special substrates, phase-change dynamic reconfiguration, and maskless rapid prototyping.
In recent years, sparse-low-rank information analysis in the polarization domain has emerged as a significant trend in polarization imaging. Such methods can effectively separate target and background information in strong scattering environments, enabling clear image reconstruction. Therefore, in-depth exploration of polarization parameter-based sparse-low-rank analysis has become a promising approach to enhance imaging performance. This study proposes an underwater polarization image restoration method based on polarization entropy and the overall purity index. The two polarization parameters are both derived from the Mueller matrix, and their integration provides a new information direction for in-depth analysis of scattered light fields. Analysis of the distribution patterns of backscattered light reveals that both polarization entropy and the overall purity index exhibit distinct low-rank characteristics. On this basis, we employ robust principal component analysis coupled with the alternating direction method of multipliers to perform low-rank and sparse decomposition on these two polarization parameters. The decomposed low-rank component represents the polarization parameters of backscattered background light. Using these parameters, we calculate the transmission map and recover the degraded images through a reconstructed underwater polarization imaging model. Experimental results demonstrate that the proposed method achieves significant improvements in image contrast and clarity across varying turbidity levels. These enhancements are effective for targets with both single and complex polarization properties.
Perimeter intrusion events are unauthorized or authorized access by humans or non-humans into a well-defined area. These events are usually related to both intrusion and non-intrusion. This paper presents a four-step iterative method (FSIM) for processing vibration signals to effectively identify events within a secure perimeter system. The first step is to de-noise the signal, and the second is to extract its features (both in the time and frequency domains). The feature extraction in the time domain involves the use of the Gramian Angular Difference Fields (GADF) algorithm, while feature extraction in the frequency domain involves the use of the Fast Fourier Transform co-generation matrix (FFTT). In the third step, each of the two algorithms will be integrated with the Convolutional Neural Network (CNN) algorithm to form two feature vectors (fvTandfvF). The two feature vectors are going to be unified to produce a single feature vector F(FT−FF). F is then normalized and used in the final feature vector Q, which is used in the multi-class Support Vector Machine (SVM) model in the fourth step to simulate the final classification. During experiments and tests, the distributed acoustic sensor (DAS) based on phase-sensitive optical time-domain reflectometry (Φ-OTDR) collects vibration signals from the fiber-optic sensing cable that surrounds the secured perimeter. The application of the FSIM shows that intrusion events can be accurately identified at about 99.29% for a processing time of approximately 0.422 s.