Accurate traffic flow forecasting plays a crucial role in alleviating road congestion and optimizing traffic management. Although numerous effective models have been proposed in existing research to predict future traffic flow, most models exhibit certain limitations in modeling spatiotemporal dependencies, especially in capturing multiscale spatiotemporal relationships. To address this, we propose a and Temporal Attention (STAIL-TA) for traffic flow prediction, which is designed for dynamic and interactive adaptive modeling of spatiotemporal features in traffic flow data. Specifically, we first design a feature augmentation layer that enhances the interaction of timebased features. Next, we introduce an interactive dynamic graph convolutional network, which uses an interactive learning strategy to simultaneously capture spatiotemporal characteristics of traffic data. Additionally, a new dynamic graph generation method is employed to design a dynamic graph convolutional block, which is capable of capturing the spatial correlations that change dynamically within the traffic network. Finally, we construct a novel temporal attention mechanism that effectively leverages local contextual information and is specifically designed for transforming numerical sequence representations. This enables the prediction model to capture the dynamic temporal dependencies of traffic flow better, thus facilitating long-term forecasting. The experimental results show that the STAILTA model improves the mean absolute error and root mean squared error on the PEMS-BAY dataset by 7.75%, 3.68% and 5.59%, 2.72% in the 15-minute and 30-minute predictions, respectively, when compared to the existing optimal baseline method, MRA-BGCN.
Culture can play a significant role in evaluating image quality. Therefore, this work considered one of the least studied cultural regions of observers, examining the impact of Central Asian culture on image quality evaluation. More specifically, it investigated how they evaluate the quality of contrast-enhanced images. It was found that observer evaluations vary and can be divided into groups. These groups may have their individual preferences for the quality of contrast-enhanced images. Therefore, the personalization factor should be incorporated into the quality evaluation of (contrast-) enhanced images. Furthermore, the results were compared with another population and differences were found in the overall outcomes of the two observer groups. The variations observed could be due to cultural differences. In addition, this study introduced the Central Asian Contrast-Enhanced Image Quality Dataset (CACEIQD). A variety of image quality metrics, including deep learning techniques, were tested on the dataset. The results indicate that the dataset is challenging and highlight an area for metric improvement. This dataset can be helpful for future research in the field of enhanced image quality evaluation.
Accurate segmentation of brain tumors is essential in the planning of neurosurgical treatments as it can significantly enhance their effectiveness. In this paper, the authors propose a modified Residual U-shaped network (ResUnet) based on multimodal fusion and a Generative Adversarial Network for multimodal brain tumor Magnetic Resonance Imaging segmentation. First, they propose a three-path structure for the encoding stage to address the issue of inadequate utilization of multimodal features, which leads to suboptimal segmentation results. The structure comprises three components: the T1 path, the T1ce path, and the fusion path combining Flair and T2 modalities. They then utilize average pooling to integrate the global information from the T1 path into both the T1ce and the fusion path, enhancing the feature fusion across different modalities and strengthening the robustness of the network. Subsequently, the features from the T1ce path and the fusion path are connected to the decoding stage through skip connections to enhance the utilization of model features and improve segmentation accuracy. Finally, the Network (DCGAN) to further enhance the accuracy of the network. They improve the loss function of the DCGAN by introducing an adaptive coefficient, which reduces the loss value in the early stages of model training and increases it in the later stages. Experimental results demonstrate that the proposed method effectively improves segmentation accuracy compared to related methods.
Purpose: Gliomas, particularly brain tumors, pose significant challenges due to their complex pathology and life-threatening potential. The goal of this study is to introduce LU-net, a novel semantic segmentation algorithm designed to enhance the diagnosis and treatment planning of gliomas. This research seeks to address the limitations of traditional classification and detection methods by improving the accuracy and robustness of tumor boundary delineation in medical images. Methods: LU-net employs a multiscale image pyramid along with a Bayesian-inference-based multiscale probability search to capture complex tumor features. The algorithm is further strengthened by integrating a Conditional Random Field model, enabling more precise segmentation. The performance of LU-net is evaluated against existing segmentation algorithms using standard metrics such as accuracy, Intersection over Union (IoU), and Dice score. Results: The experimental results demonstrate that LU-net outperforms current segmentation algorithms in terms of both accuracy and robustness. Specifically, LU-net achieves an accuracy of 0.9953, an IoU of 0.667, and a Dice score of 0.566, effectively addressing the pathological heterogeneity and invasiveness of gliomas. These results highlight LU-net's superior ability to delineate tumor boundaries and improve diagnostic accuracy. Conclusion: LU-net sets a new benchmark in glioma lesion detection, offering a more effective approach for brain tumor segmentation. By improving the accuracy, reliability, and interpretability of brain tumor boundary delineation, LU-net enhances diagnostic and treatment strategies, providing significant benefits to patients, clinicians, and healthcare providers. Overall, this work marks a significant contribution to the field of medical imaging and glioma diagnosis. c 2026 Society for Imaging Science and
In the present digital age, where visual stimuli dominate our experiences, image quality plays an important role in determining user satisfaction. In this study, we propose a new direction for advancing the field by focusing on personalized image quality assessment, which can benefit a wide range of industries from streaming services and photography postprocessing algorithms to more specialized domains with different image quality requirements, such as medical imaging. We focus on exploring individual preferences, specifically in the context of contrast, one of the key attributes influencing image quality. We employ a twin neural network to predict individual contrast preference, which is particularly effective for enhancing contrast adjustment for observers with distinct and consistent preferences that deviate from the average. We also introduce an intermediate step toward personalization-a faster and less computationally intensive approach for incorporating observer awareness into general quality assessment models. Our findings highlight the importance of accounting for individual preferences and the positive impact of incorporating them into image processing and quality assessment methodologies.
Time-resolved imaging requires image sensors with high-speed modulation and a thick sensing layer for high sensitivity to near-infrared light and high-energy photon and particle detection. A solution is a lock-in pixel based on a silicon on insulator (SOI) with a fully depleted thick substrate; however, such pixels face challenges from degraded modulation contrast due to high parasitic light sensitivity (PLS) in the floating diffusion region. This work presents a pixel architecture that suppresses the PLS while preserving fast charge transfer by introducing a shallow buried channel and intermediate gates that enhance the lateral electric field. A prototype four-tap pixel array was fabricated using 0.2 & micro;m SOI technology and experimentally characterized under 940 nm pulsed laser illumination. The proposed structure achieved a modulation contrast of 92.5% and PLS below 1.2% under 40 ns gate-pulse operation, maintaining stable performance up to 35,000 electrons per pulse, demonstrating suitability for X-ray and electron-beam imaging.
An interior graphic design method based on the graphic generation network and the dual-branch Transformer is proposed aimed at the problems of insufficient layout rationality and difficulty in coordinating and optimizing style and function in the automation of interior graphic design, especially the low manual operation efficiency of computer-aided design (CAD) systems and the poor engineering adaptability of existing automation methods. The topological structure of functional areas and spatial connections is constructed through the graph generation network. The layout is incrementally generated by combining the improved breadth-first search algorithm, and the semantic prediction network is introduced to achieve the collaborative optimization of geometry and semantics. The dual-branch Transformer processes geometric topology and functional semantics respectively, optimizes detail design by using the cross-modal attention mechanism, and dynamically adjusts the feature fusion weights. Experiments show that this method achieves an average intersection and union ratio of 84.03% and a pixel error of 4.84% in layout generation quality, with a processing speed of 0.09 s per scheme, meeting the real-time interaction requirements of CAD tools. The generated graphic design scheme achieved a Peak Signal-to-Noise Ratio of 34.17 dB and a structural similarity of 0.91 in visual quality evaluation, showing a high degree of consistency with the professional design scheme, indicating that the generated scheme has high clarity and structural rationality. Compared with the existing methods, this method demonstrates significant advantages in terms of generation efficiency, layout rationality, and design diversity. Compared with the real label drawings, the generated results are close to the actual design requirements in terms of space utilization and layout consistency. This research, through the dual-branch collaborative modeling of geometry and semantics, has significantly enhanced the automation level of interior graphic design and the practical value of generating solutions in CAD integrated scenarios.
Selecting lattice networks to achieve specific tailored material properties has traditionally been a daunting task. Unit cell selection is a "heuristic-based" methodology, which is timeconsuming and rarely leads to an optimal solution. A new approach to metamaterial design methodology encompassing quantitative unit cell selection and optimization that is based on baseline geometry is presented. To achieve this new design roadmap, a real-world case is used for utilizing metamaterials to design an optical bench from Aluminum 6061 T6 equivalent (Al6061 RAM2), achieving 2-micron surface deformation and a 10% mass penalty relative to Beryllium I-220H of diametrical surface-level deformation. The primary goal is to design specific beryllium-like mechanical properties without the added manufacturing challenges, lead time, and cost of Beryllium I-220H. Quantitative lattice selection methodology is considered in which a lattice network design is developed to reduce the structure weight while still maintaining overall resistance to deformation when a thermal load is applied to the optical bench. The result is a quantitative design process that can produce metamaterial geometry tailored to specific material properties in less than 100 days including manufacturing. c 2026 Society for Imaging Science and Technology.
To address the issue of low accuracy in 3D modeling of images captured by unmanned aerial vehicles (UAVs), the authors propose an enhanced 3D reconstruction model for a mountain in Yuanmou by employing an improved Structure from Motion-Multiview Stereo (SFM-MVS) algorithm. In the process of converting 2D into 3D data, the key challenges lie in feature point extraction and matching. The authors introduce an algorithm for optimizing Speeded Up Robust Features (SURF) by combining the SURF descriptor operator with a fast feature point algorithm. The use of the Laplace operator to refine the extraction of weighted feature points along with integration into the robust SURF descriptor simultaneously improves both matching speed and accuracy. This solution mitigates issues related to excessive image data and the low accuracy and efficiency of 3D reconstruction models in UAV applications. Experimental results demonstrate that the proposed method extracts more feature points, increases matching speed, and significantly enhances accuracy compared to both the original SURF and traditional Scale-Invariant Feature Transform algorithms. When compared to the unoptimized SFM-MVS algorithm, the accuracy of the optimized 3D reconstruction using the SURF-based algorithm improves by approximately 44.68%, with a 30% increase in processing speed. Additionally, to evaluate UAV path planning performance on complex terrain, the authors first employ the optimized SURF-based 3D reconstruction method to precisely reconstruct the terrain map of the target area. This method improves both the accuracy and efficiency of 3D terrain reconstruction, providing high-precision data for subsequent path planning algorithm performance testing. Subsequently, three classical path planning algorithms-Ant Colony Optimization, A*, and Rapidly exploring Random Tree-are selected for comparative analysis of UAV path planning capabilities on complex terrain.
In this article, the authors introduce a versatile framework and rendering techniques for the construction of cost-effective and easy-to-assemble multi-view 360 degrees displays. The proposed methodology allows for the creation of displays featuring high-resolution and full-color images across a variable range of viewing angles. This framework extends the existing technology, which utilizes a parallax approach using rotating screens. A variety of display configurations achievable through the proposed framework are explored, including a multi-display strategy to mitigate flicker, as well as a flat design aimed at delivering a see-through experience. Additionally, the rendering workflow used to generate multi-view 360 degrees content is introduced. Finally, results are presented in the form of photographs captured from varying angles along with an analysis of the displayed content derived from two constructed prototypes: the first exemplifies a cylindrical design with dual displays while the second showcases a see-through flat-based design. c 2026 Society for Imaging Science and Technology.
This paper presents a comprehensive experimental evaluation of spectral reconstruction methods in multispectral imaging systems, focusing on two multispectral camera technologies with differing spectral characteristics: spectral filter array and filter wheel. These systems were assessed under a controlled LED-based illumination setup. A range of reconstruction methods, encompassing both model-based and training-based approaches, were analyzed in their baseline forms as well as in adaptive configurations, which select optimal local training subsets based on spectral reflectance or camera response similarity. Experiments were conducted using a custombuilt imaging setup and two well-characterized spectral reflectance datasets: the standard Munsell and the Munsell Student Color sets. Results demonstrate that training-based methods significantly outperform model-based methods in both spectral and colorimetric accuracy. Adaptive dataset selection further enhances performance in many cases, particularly for the SpectroCam filter wheel camera. The influence of illumination on reconstruction accuracy is also examined, revealing that model-based methods are especially sensitive to the spectral power distribution of the light source. These findings offer practical and technical guidance for the design and calibration of multispectral imaging systems aimed at achieving high-accuracy spectral recovery.
The maintenance of critical transportation infrastructure such as roads, tunnels, and bridges depends heavily on timely and accurate detection of structural damage. Among various types of surface defects, concrete cracks are the most common and dangerous. However, traditional manual inspection methods are inefficient, labor-intensive, and pose safety risks. In this paper, the authors propose UC-Net, a novel semantic segmentation network specifically designed for concrete crack detection. The UC-Net builds upon the U-Net architecture and introduces a Coordinate Attention mechanism into skip connections, enabling the model to better capture long, narrow, and spatially scattered crack features while suppressing irrelevant background noise. This design addresses two key challenges: the small spatial proportion of cracks in high-resolution images and the interference from complex textures and lighting conditions. To validate the effectiveness of this approach, the authors conducted extensive experiments on a publicly available crack dataset and real-world images. Compared with existing CNN- and attention-based networks, UC-Net achieves superior performance, with improvements in IOU (up to 91.78%) and accuracy (89.33%). The results confirm that UC-Net provides a lightweight yet accurate solution for fine-grained crack segmentation, and it can serve as a practical tool in real-world infrastructure monitoring scenarios.
The technology of generating tactile data from visual modalities holds significant importance in cutting-edge fields such as tactile rendering, virtual reality, and robotics. This technology effectively bypasses the cumbersome process of manual tactile data collection and overcomes the limitations inherent in physical contact, thereby opening new avenues for advancement in related fields. However, current methods suffer from notable drawbacks: they struggle to ensure consistent and reliable results when generating tactile data across different categories, which greatly restricts their practical applications. To address this challenging problem, the authors have developed the T-CGAN cross-modal generation framework. Based on the FrictGAN architecture, this framework innovatively introduces an image category conditional constraint mechanism and texture feature extraction combined with the L1 loss function to precisely regulate the generation process and ensure high-quality output. Specifically, the framework can generate spectrograms of friction coefficient signals from fabric texture images and then convert these spectrograms into one-dimensional friction coefficient signals using the Griffin-Lim algorithm. During the research, the authors employed root mean square error and mean absolute error metrics to quantitatively analyze the differences among generated spectrograms, reconstructed signals, and their corresponding ground truths and conducted a comprehensive comparison with existing methods. Extensive experimental results demonstrate that this method significantly outperforms existing techniques in terms of both accuracy and stability, providing a superior solution for the field of tactile data generation.
Graph neural networks (GNNs) have emerged as powerful tools for news text classification by explicitly modeling inter-document relationships and complex semantic dependencies. This paper presents the first comprehensive survey specifically examining GNN methods tailored for news text classification, addressing the unique challenges that distinguish news data from general text: temporal dynamics of rapidly evolving events, multi-dimensional complexity across political-economic-social aspects, dense entity networks, and credibility verification requirements. The authors systematically analyze GNN-based approaches from three core dimensions: (1) graph construction methods, from word co-occurrence and syntactic-semantic relationships to news-specific temporal-aware and event-centric graphs; (2) feature extraction mechanisms leveraging both node and edge characteristics; and (3) four major GNN architectures-GCN, GAT, GGNN, and GTN-evaluating their specific advantages for news classification tasks. The survey identifies critical challenges including computational scalability for real-time news streams, dynamic graph adaptation, multi-modal integration, and interpretability requirements. This comprehensive review provides researchers and practitioners with a systematic understanding of how GNN architectures can be adapted to address the distinctive requirements of news text classification, establishing foundation for future developments in this rapidly evolving field.
This paper presents an adaptive method for extracting fabric pattern templates, focusing on efficiently and accurately digitizing textural features in traditional fabric images. Using Segment Anything Model 2 (SAM2) automatic mask generation as the core, this study precisely segments color blocks in fabric images, providing high-quality data for further processing. The method employs a multistep strategy. First, color quantization and bilateral filtering reduce image complexity, remove noise, and enhance edges. Second, advanced edge detection algorithms identify prominent edges to assist SAM2 segmentation, ensuring accuracy and reliability. Finally, masks generated by SAM2 are classified and merged based on their covered colors in the original images, producing clear pattern templates. This method is validated with numerous real fabric images, and it shows strong adaptability and efficiency in extracting color templates. It provides robust support for digital preservation of traditional fabric patterns and opens up opportunities for innovative applications and heritage development, marking a significant advancement in this field.
Knowledge graphs play a critical role in intelligent systems, but they face persistent challenges of incomplete data acquisition, noisy information, and inefficient inference under dynamic updates. To address these issues, the authors propose a graph-embedding-based framework that integrates three novel components: (1) a neighborhood-enhanced embedding module that captures richer structural semantics, (2) an inference optimization mechanism based on contextual consistency and confidence reweighting, and (3) a dynamic update strategy for efficient incremental learning. Extensive experiments on FB15k-237, WN18RR, and MedKG show clear improvements over state-of-the-art baselines. The proposed framework achieves Mean Reciprocal Rank gains of 8-15% and Hits@10 gains of 3-6%, demonstrating substantial accuracy improvements in link prediction. On dynamic update tasks, the proposed method maintains almost identical accuracy to full retraining (AUC difference < 0.2%) while achieving a 7.7-fold reduction in update time. These results verify that the proposed framework significantly enhances both the effectiveness and efficiency of knowledge graph reasoning.
High-visibility watermarks with strong visual saliency are particularly vulnerable to removal by deep learning models. Existing adversarial-perturbation-based protection methods for visible watermarks often suffer from poor perturbation stability and significant image quality degradation in high-visibility scenarios. To address these challenges, the authors propose a cascaded coarse-to-fine framework that generates an adversarial perturbation High-visibility Watermark Vaccine (HWV), specifically aimed at protecting high-visibility watermarks. They establish a watermark-driven perturbation generation model and design cascaded loss functions for coarse and fine stages to guide the multi-phase search for a globally optimal adversarial solution. In the coarse stage, a composite loss function is constructed to achieve robust protection of high-visibility watermarks while in the fine stage, a perturbation minimization objective is introduced to mitigate the impact of perturbations on image quality. Moreover, the authors propose a novel gradient normalization equation combined with a dynamic momentum update strategy to adaptively optimize the perturbation step size, accelerating convergence toward the global optimum of the loss function. Experimental results on the CLWD dataset demonstrate that the proposed method effectively prevents removal attacks targeting high-visibility watermarked images. Furthermore, compared to conventional single-stage loss methods, this method significantly improves the image quality of perturbed watermark images, achieving a Peak Signal-to-Noise Ratio greater than 44 dB. This work provides a novel perspective for enhancing digital image copyright protection against deep-learning-based attacks.
High dynamic range (HDR) imaging techniques effectively enhance image quality in driving scenarios. However, HDR image synthesis in automotive cameras remains challenging due to complex conditions such as high-contrast scenes, low-light environments, and vehicle motion. Automotive cameras typically optimize image quality by adjusting exposure time to control the duration of light capture and tuning analog gain to amplify the sensitivity of the imaging sensor. The study aims to identify the optimal parameter combinations of exposure times and analog gain settings for HDR image synthesis in automotive cameras. The authors acquire images at three frames captured under different exposure time and analog gain combinations and use HDR image synthesis techniques to simulate HDR images. A composite quality evaluation method was established based on four dimensions: tone range, tone levels, contrast, and signal-to-noise ratio. Quantitative analysis revealed that the best HDR image quality was achieved when the exposure time ranged from 5 ms to 20 ms and the analog gain ranged from 1 & times; to 2 & times;. The proposed HDR image synthesis strategy demonstrates significant practical value for automotive vision systems, improving image quality and providing more accurate and reliable visual information for autonomous driving and advanced driver-assistance systems, enhancing driving safety and user experience.
The reconstruction of 3D digital humans has extensive practicality in various fields at present, including virtual reality, augmented reality, film and game production, visual effects, and so on. Recently, 3D Gaussian Splatting (3DGS) has achieved high-quality novel view synthesis and fast rendering and can also be applied to human reconstruction. However, it is very difficult to directly edit the reconstructed human model. Therefore, it is necessary to extract the geometric shape of the human body from the Gaussian model. However, since the optimized Gaussian model is unorganized, it produces a noisy and unrealistic surface. Thus, accurately extracting the mesh from the Gaussian model remains a challenge. To address this issue, the authors propose GSHumanMesh, which extracts a mesh from Gaussians with surface alignment for human mesh reconstruction. They utilize a set of multi-view human images as the input and obtain the Gaussian model via 3DGS. They also propose an improved surface regularization term based on SuGaR and 2DGS to enforce Gaussians to be surface-aligned. Then they extract an accurate mesh from the Gaussians using Poisson reconstruction. Additionally, the authors propose post-processing to refine the reconstructed mesh and make it smoother. The method was tested on the THUman4.0 and DualGS datasets. Compared with other stateof-the-art methods based on 3DGS, the resulting reconstruction is smoother, more accurate, and shows more intricate details.
Register accuracy is an important indicator for measuring the quality of multicolor gravure printing. However, the accuracy of the photoelectric sensor register error detection does not meet the register control requirements for light-colored printed electronic products. To address this issue, the authors propose a high-precision register error detection method based on machine vision for circular matrix color marks. First, image denoising is carried out based on the characteristics of circular matrix color marks. Then a combined coarse-fine detection method is employed, and an improved Canny operator is used to detect pixel-level edge points in color marks. The eight-neighborhood edge tracking is performed by connecting the edges through morphological dilation processing. Then, based on pixel-level edges, the Zernike moment method is used to obtain subpixel-level edge points. Finally, the least squares method is utilized to fit the color mark center based on subpixel edges, and the register errors in both horizontal and vertical directions are obtained based on the distances between each color mark and the reference color mark center. Experimental results show that the proposed register error detection method based on matrix color marks achieves an average register error of 0.026 pixels. This method not only enhances the accuracy of register error detection but also demonstrates stability against noise and varying lighting conditions, meeting the demand for register error detection in complex gravure printing equipment.
