最新刊期
卷 34 ， 期 18 ， 2026
- 摘要：In order to achieve accurate identification of aquatic bacteria， this study employed transmission spectroscopy combined with a generative adversarial network （GAN）-based data augmentation approach to classify and identify five typical aquatic bacterial species. Addressing challenges such as the high cost of acquiring transmission spectral samples， susceptibility to instrument performance variations， and insufficient model generalization under small-sample conditions， we proposed a spectral data augmentation strategy based on a dynamic filtering generative adversarial network （DFGAN） and constructed a convolutional neural network （CNN） classification model. Initially， 600 transmission spectra of the five bacterial species were collected across different cultivation periods and concentration gradients to form the original dataset. Subsequently， a deep convolutional generative adversarial network （DCGAN） and the proposed DFGAN were implemented for data augmentation， with the quality of generated spectra evaluated from three dimensions： visual morphology， statistical distribution similarity， and feature space diversity. Finally， comparative analyses were conducted on the recognition performance of three classifiers： support vector machine （SVM）， K-nearest neighbor （KNN）， and CNN. The results indicated that the quality of spectra generated by DFGAN was significantly superior to that of DCGAN. The aquatic bacteria identification model established by the DFGAN-CNN approach exhibited optimal performance， achieving an accuracy of 95.00%， precision of 95.45%， recall rate of 95.00%， and F1-score of 95.00%， thus demonstrating the highest accuracy and robustness. This method not only provides a novel approach for the rapid detection of aquatic bacteria but also offers technical references for addressing small-sample challenges in the field of spectral analysis.关键词：transmission spectrum;data augmentation;waterborne bacteria;recognition;dynamic filtering generative adversarial network4|0|0更新时间：2026-10-08
- 摘要：To address the issues of insufficient accuracy and frequent violations of physical constraints during the interpolation and reconstruction of measured data in Polarized Bidirectional Reflectance Distribution Function （pBRDF） databases—which previously degraded the accuracy and credibility of polarization rendering results—this paper proposed a physics-constrained pBRDF reconstruction method tailored for polarization rendering. First， the pBRDF data was mapped into cylindrical coordinates to eliminate parameter singularities and boundary discontinuities. Second， an initial Mueller matrix was obtained via a local linear interpolation method based on periodic distance metrics. Subsequently， reciprocity symmetrization， Cloude semi-definite eigenvalue truncation， and energy conservation projection were applied sequentially to ensure the physical realizability of the reconstructed results. Finally， the reconstructed data was integrated into a polarized ray tracing pipeline for rendering validation. Experimental results based on the KAIST dataset demonstrate that the proposed method reduces the storage size of single-material data from 912 MB to 650 MB. It achieves a zero physical violation rate for both Gold and Purple Silicone materials. Compared with the TD method， the proposed approach improves the SSIM of the Degree of Polarization images for Gold and Purple Silicone materials by 4.3% and 0.8%， respectively， and the SSIM of the Angle of Polarization images by 4.5% and 5.0%， respectively. By balancing reconstruction accuracy， physical correctness， and storage efficiency， this method provides effective support for the high-precision reconstruction of pBRDF measurement data and realistic polarization rendering.关键词：polarization bidirectional reflectance distribution function;Mueller matrix;cylindrical coordinate representation;physically constrained interpolation;polarization rendering3|0|0更新时间：2026-10-08
- 摘要：Snow plays an important role in the Earth's energy balance， hydrological cycle， and climate system， and its reflection characteristics are fundamental to characterizing radiative transfer processes at the land surface.Snow also serves as one of the primary ground feature backgrounds on winter battlefields. In natural environments， snow surfaces are often shaped by wind and underlying terrain， forming macroscopic structures such as snow ridges， which can alter photon propagation paths and scattering directions， thereby affecting the polarized reflection of snow. However， the effects of such macroscopic structures on snow polarized reflection and the applicability of existing polarized reflectance models remain poorly understood. To address this issue， this study investigated both flat snow surfaces and snow-ridge surfaces. Three observation scenarios were designed by changing the relative orientation between the solar incidence direction and the ridge direction： a flat snow surface， ridges parallel to the incidence direction， and ridges perpendicular to the incidence direction. A multi-angular polarimetric observation system was used to measure the hemispherical distribution of the polarized reflectance factor under these three scenarios. Based on the measurements， the performance of two widely used polarimetric reflectance models， the Litvinov model and the Nadal–Bréon model， was evaluated. The results show that for flat snow surfaces and snow ridges parallel to the incident direction， the bidirectional polarized reflectance factor exhibits relatively regular hemispherical distributions， primarily characterized by a pronounced enhancement in the forward-scattering direction. When the snow ridges were oriented perpendicular to the illumination direction， the polarized reflectance in the forward scattering direction was substantially weakened， while the polarization was improved in part of the backward scattering direction. Model evaluation indicates that both the Litvinov and Nadal-Breon models achieve high simulation accuracy for flat snow surfaces， with R² values of 0.86 and 0.84， respectively， and relative root mean square errors （RRMSEs） below 29%. However， model performance significantly decreases for ridged snow surfaces. For snow ridges parallel to the incident direction， R² values decrease to 0.73 and 0.64， with RRMSEs of approximately 39%. When the snow ridges were perpendicular to the illumination direction， the weakened forward-scattering and the improved backward-scattering polarization caused the angular distribution of polarized reflectance to deviate from the assumptions of the models. Consequently， both models failed to accurately reproduce the angular distribution of polarized reflectance and exhibited their poorest performance， with R² values of 0.43 and 0.44， respectively and RRMSEs of approximately 50%. Overall， this study reveals the significant influence of snow surface structure on polarized reflectance characteristics and highlights the limitations of existing polarized reflection models in representing ridged snow surfaces. These findings indicate uncertainties in the application of polarimetric remote sensing under complex surface conditions and provide a basis for improving and developing snow polarized reflectance models.关键词：multi-angular polarized reflectance;snow cover;snow ridges;polarized reflectance models3|0|0更新时间：2026-10-08
Modern Applied Optics
- 摘要：Existing seamless steel pipe surface defect detection methods depend on labeled samples. They also rely on a single sensing modality and incur high hardware costs. They further struggle to simultaneously satisfy detection efficiency and accuracy requirements under industrial conditions. To address these limitations， this paper proposed a defect detection method. The method is based on line laser scanning with 2D and 3D fusion. A dedicated optical path was designed to acquire surface data. This design achieved pixel-level alignment of grayscale intensity and depth information. A relative ratio method was proposed to avoid calibrating surface reflectance under complex illumination. A parametric reference template generation mechanism was further established. This mechanism enabled automatic recognition of multiple pipe diameters and batch-level parameter reuse. A hierarchical relative-position computation strategy was proposed for complex inspection environments. This strategy suppressed measurement disturbances induced by conveyor belt vibration. A dual-comparison algorithm was developed， integrating grayscale texture analysis with geometric deviation assessment. This algorithm identified typical surface defects， including grooves， pits， perforations， and scale. Experimental results demonstrate that the system achieves a comprehensive detection accuracy of 99.40%. The system reliably detects surface defects as small as 0.50 mm. This performance is maintained under a pipe transfer speed of 1.2 m/s. The proposed system realizes full circumferential coverage inspection of the steel pipe outer surface. It satisfies the accuracy and real-time requirements of industrial online inspection.关键词：steel pipe surface;defect detection;line laser;three-dimensional vision;pipe diameter measurement4|0|0更新时间：2026-10-08
- 摘要：Glass aspheric cylindrical lens arrays offer low aberration， high collimation efficiency， excellent beam homogenization capability， and a high degree of optical integration， showing considerable potential for applications in high-power laser systems and optical communication coupling. However， their fabrication remains challenging because of the difficulty in machining high-precision array molds and the limited controllability of form accuracy during precision glass molding. In this study， a simulation model of the cylindrical lens array was established to optimize the mold design and predict form errors， thereby enabling mold pre-compensation. A complete process chain for fabricating microstructured array molds for low-melting-point glass was proposed， including mold substrate selection， heat treatment of the Ni-P electroplated layer， single-point diamond planing， and Ti-DLC coating to enhance mold hardness and wear resistance. Precision molding experiments were subsequently conducted using a multi-step heating and annealing strategy combined with segmented control of the molding rate. The experimental results show that the fabricated aspheric cylindrical surfaces achieved an average areal surface roughness of Sa 1.553 nm and an average form accuracy of PV 0.208 μm. The molded glass aspheric cylindrical lens arrays exhibited a form accuracy of PV≤0.3 μm and a surface roughness of Sa≤3 nm. The proposed mold pre-compensation method and integrated manufacturing process provide an effective technical reference for the mass production of low-melting-point glass optical components with microstructured array features.关键词：aspheric cylindrical lens array;glass lens;pre-compensation;mold fabrication;array precision glass molding2|0|0更新时间：2026-10-08
- 摘要：Aiming at the common problems of insufficient measurement accuracy and poor high-temperature stability of traditional force sensors in dynamic force monitoring under high-temperature conditions， a novel piezoelectric force sensor suitable for medium and high-temperature environments ranging from room temperature to 350 ℃ and featuring excellent dynamic response performance was developed. Meanwhile， a sensor design method integrating single-factor sensitivity analysis， nonlinear regression modeling， and parameter optimization based on the Love Evolutionary Algorithm （LEA） was proposed. A lumped-parameter dynamic model of the sensor considering mass， stiffness and damping was established via mechanism modeling. Five core parameters affecting the natural frequency were screened out using the control variable method， and a high-precision nonlinear regression model with interaction terms and higher-order terms was constructed in combination with stepwise regression. Taking the maximization of natural frequency as the optimization objective， the structural optimization of the sensor was completed using the Love Evolutionary Algorithm. A high-temperature dynamic performance test system was built to experimentally evaluate the dynamic characteristics of the developed sensor at elevated temperatures. The results show that the nonlinearity of the sensor is 1.07%， and the repeatability error is 1.37%. Within the preset temperature range， its dynamic sensitivity remains stable above 30.33 pC/N， and the temperature drift coefficient is only -0.014%/℃. The research results can provide a theoretical reference for the structural design and parameter optimization of piezoelectric force sensors used in high-temperature working conditions.关键词：high temperature;piezoelectric force sensor;love evolution algorithm;dynamic force detection;structural optimization2|0|0更新时间：2026-10-08
Micro\/Nano Technology and Fine Mechanics
- 摘要：To address the limitations of traditional UAV monocular vision-based ground target localization methods， which relied on prior information of the target or scene and suffered from insufficient positioning accuracy， a precise ground target localization method using up to two observation points was proposed. First， considering factors that affected positioning accuracy—including UAV pose and gimbal attitude variations， deviation of the gimbal rotation center from the UAV center of gravity， and deviation of the imaging system optical center from the gimbal rotation center—a projection model of ground targets in the aircraft geographic coordinate system was established， enhancing the model's generality for UAV-based ground target observation. Subsequently， based on ground flatness， single-observation-point and dual-observation-point ground target localization algorithms were designed， efficiently solving target position coordinates through geometric equations. Using the Phantom 4-RTK UAV as an example， Monte Carlo simulations and actual experiments were conducted to analyze the influencing factors and effects on UAV ground target localization accuracy， thereby validating the effectiveness of the proposed method. Simulation results indicated that the maximum positioning accuracy of the Phantom 4-RTK UAV for ground targets reached the decimeter level， with measurement accuracy of altitude and gimbal attitude angle being the dominant factors affecting positioning error. Positioning stability was ensured when the intersection angle between the dual observation points and the target exceeded 60°. Experimental results demonstrate that at UAV altitudes of 10-20 meters， the proposed method achieves a positioning error of 0.34 m for flat ground using a single observation point， representing a 60% reduction compared to existing methods. For non-flat ground using dual observation points， the positioning error is 0.47 m， a 92.27% reduction relative to existing methods. The proposed method achieves precise target localization in both flat and non-flat ground scenarios， demonstrating strong application potential.关键词：UAV;target localization;monocular vision;dual observation points2|0|0更新时间：2026-10-08
- 摘要：Ghost imaging is inevitably affected by complex noise in practical scenarios， making it difficult to simultaneously achieve structural recovery and noise suppression. To address this problem， this paper proposed a two-stage snapshot ghost imaging reconstruction framework combining Tikhonov regularization with deep denoising. In the first stage， Tikhonov regularization was employed to solve the inverse problem of ghost imaging， where the regularization constraint suppressed ill-posedness to obtain a stable initial reconstruction with well-preserved structural information. In the second stage， a Noise2VST module incorporating a physics-based data-fidelity constraint was developed as an adaptive deep denoising module through joint optimization. The residual complex noise was first mapped into an approximately additive white Gaussian noise distribution via a variance-stabilizing transform. Adaptive denoising was then performed under the joint constraints of self-supervision and a physics-based data-fidelity constraint， effectively removing residual noise while preserving image details and edge structures. Experimental results demonstrate that the proposed method consistently achieves favorable reconstruction performance under different sampling conditions.Under high-sampling conditions， compared with the Tikhonov method， the proposed method improves PSNR by 4.68 dB and SSIM by 0.375. Under lower-sampling conditions， it maintains competitive visual quality and quantitative performance， demonstrating its effectiveness and robustness.关键词：ghost imaging;Tikhonov regularization;Variance-Stabilizing Transform（ VST）;image reconstruction2|0|0更新时间：2026-10-08
- 摘要：To address the sensitivity to initial contours and poor noise robustness in traditional level set segmentation algorithms， this paper proposed a segmentation model based on global Gaussian distribution and statistical norm metric. By fusing edge information with global region information， the model achieved accurate target boundary extraction and enhanced robustness against initial contour placement and noise interference. First， a novel global energy term was designed as the data term using a global information fusion strategy. This term， combined with a regularization term， drove the initial contour toward target boundaries for precise region segmentation. Second， a global Gaussian distribution was defined to fit the energy using the level set function， with global mean and variance treated as variables. Finally， energy minimization was achieved through alternating level set evolution and estimation of global intensity mean and variance during iterations. Experiments on synthetic， real， and natural images demonstrate that the proposed model exhibits excellent segmentation performance for Gaussian noise， speckle noise， and salt-and-pepper noise. Furthermore， the model outperforms other classical active contour models in terms of segmentation accuracy and computational efficiency. The proposed model improves segmentation accuracy for high-noise images， achieves effective target boundary extraction， and demonstrates superior performance in both segmentation precision and computational efficiency.关键词：image segmentation;active contours;statistical norm;level set;Gaussian distribution2|0|0更新时间：2026-10-08
- 摘要：Aiming at the challenges of large scale spans， complex roof textures， and background interference in high-resolution remote sensing imagery， which caused traditional models to suffer from internal holes when extracting large buildings and boundary geometric feature loss for small targets， a dual-stream network architecture integrating multi-scale local features and global spatial representations was proposed. A hybrid network based on the fully symmetric U-Net topology was constructed. In the encoding stage， a hierarchical feature pyramid was designed to extract multi-resolution local texture information， and a Vision Transformer module was embedded at the deepest layer （1/16 of the original spatial resolution） to accomplish long-range dependency modeling using a multi-head self-attention mechanism. In the decoding stage， an adaptive dual attention recalibration mechanism was introduced， utilizing a cascaded Convolutional Block Attention Module （CBAM） to suppress noise in features transferred via skip connections from both spatial and channel dimensions， while a cross-attention module was fused to align multi-scale features. At the network's end， a multi-task output branch with parallel segmentation and boundary prediction was constructed， which was jointly optimized by using a boundary loss function. On the WHU aerial imagery dataset， the proposed model achieves an intersection over union （IoU） of 89.48%， an F1 score of 94.45%， a precision of 90.12%， and a recall of 99.21%. Compared with the baseline U-Net architecture， the IoU and F1 score are improved by 13.95% and 14.05%， respectively. On the ISPRS Potsdam dataset with complex backgrounds， the model's recall rate remains stable at 90.48%. Ablation experiments demonstrate that the introduction of global self-attention module increases IoU by 6.67%， effectively mitigating internal feature loss of the targets. With controllable computational overhead， the proposed network architecture effectively extracts the spatial topology and local contours of multi-scale buildings， overcoming the limitation of the local receptive field in two-dimensional convolutional models. It exhibits significant practical deployment value in engineering applications such as urban spatial evolution monitoring and geographic basic database updating.关键词：building extraction;semantic segmentation;multi-head self-attention mechanism;feature recalibration;multi-task learning2|0|0更新时间：2026-10-08
- 摘要：Cross-domain few-shot hyperspectral image classification faced the dual challenges of domain shift and label scarcity. Existing methods predominantly relied on pure visual features or adopted fixed text-visual alignment strategies， which struggled to capture fine-grained spectral discriminative information and overlooked the quality differences among target domain samples. To address these issues， a text-enhanced dynamic contrastive and adaptive prototype guidance classification method was proposed. Firstly， a dual-branch multi-scale network extracted spatial-spectral visual features from image patches， while a CLIP text encoder was introduced to obtain class-level semantic representations serving as cross-domain stable anchors. On this basis， a text-enhanced dynamic contrastive learning strategy was designed， which dynamically adjusted the weights of hard negative samples according to the degree of visual confusion， thereby enhancing the discriminative ability for fine-grained categories. Meanwhile， an adaptive prototype guidance loss was constructed， which adaptively fused text anchors and source-domain visual neighbors based on the discriminative salience of target-domain prototypes in the source feature space， indirectly improving the reliability of target-domain prototypes through gradient back-propagation. Cross-domain experiments are conducted with Chikusei as the source domain and Indian Pines， Houston， and Salinas as the target domains. The overall classification accuracy of the proposed method reaches 81.22%， 76.82%， and 93.32%， respectively， representing an improvement of approximately 2%-5% over existing mainstream cross-domain few-shot classification methods. Through the effective synergy of dynamic contrastive learning and adaptive prototype guidance， the proposed method significantly improves the classification accuracy and stability in cross-domain few-shot scenarios.关键词：hyperspectral image classification;cross-domain few-shot learning;dynamic contrastive learning;adaptive prototype guidance;cross-modal alignment3|0|0更新时间：2026-10-08
- 摘要：To address insufficient image brightness， blurred edges， and poor distinguishability of key structures caused by inadequate tunnel illumination， uneven supplementary lighting， and local shadows in vehicle-mounted railway fastener detection， a Multi-Scale Structure-Aware Network （MSA-Net） was proposed for low-light image enhancement. Built upon the Self-Calibrated Illumination （SCI） network， MSA-Net employed a Multi-Scale Feature Extraction Module （MSFE） to extract illumination， texture， and edge information under different receptive fields， a Structure-Sensing Attention Module （SSA） to strengthen the responses of bolt contours， clip textures， and fastener boundaries， and an Adaptive Residual Fusion Module （ARF） to dynamically adjust the enhancement intensity according to regional degradation levels. Ablation studies， algorithm comparisons， and downstream detection experiments were conducted on a low-light railway fastener image dataset collected from the Fuzhou Binhai Express Line. The results show that MSA-Net achieves a standard deviation， average gradient， information entropy， Natural Image Quality Evaluator， and Blind/Referenceless Image Spatial Quality Evaluator of 46.39， 30.49， 4.88， 21.36， and 26.34， respectively. The model contains 0.016 M parameters， requires 9.782 GFLOPs， and achieves an enhancement speed of 302.39 FPS. For the YOLOv8 detector trained on normal-light images， the Precision， Recall， mAP@0.5， and mAP@0.5：0.95 obtained using MSA-Net-enhanced images reach 86.10%， 83.60%， 85.80%， and 59.00%， respectively， with an end-to-end processing speed of 99.02 FPS. For the detector trained on original low-light images， the corresponding four metrics reach 88.10%， 85.40%， 88.60%， and 61.80%， respectively. Experiments on two public low-light image datasets further demonstrate the adaptability of the proposed method to different scenes. MSA-Net improves low-light railway fastener image quality and structural distinguishability while maintaining real-time processing efficiency， thereby providing stable image inputs for vehicle-mounted railway fastener detection.关键词：low-light image enhancement;railway fasteners;multi-scale feature extraction;structure-sensing attention;adaptive residual fusion2|0|0更新时间：2026-10-08
Information Sciences
- Address：No.3888 Dong Nanhu Road, Changchun, Jilin, China Postal code：130033
- Tel：0431-86176855 Email：gxjmgc@ciomp.ac.cn
- Technical support is provided by Beijing Founder electronics co., LTD 吉ICP备11002662号-17
京公网安备11010802024621 - It is recommended to read the content of this site in Chrome&IE9+. Please switch to extreme mode in browser 360.
- Cookies We use cookies to help provide and enhance our service and tailor content. By continuing, you agree to the use of cookies.
0