Computer tomography (CT) technology is widely used in non-destructive testing. To improve detection speed, fewer sampling angles are usually used, but this poses a challenge to the accuracy of image reconstruction. Deep learning is currently the latest technology in the field of fault imaging, but due to the lack of emphasis on reconstructing detailed information, the performance of existing methods is poor. Therefore, this article proposes a dual-view projection 3D CT reconstruction network based on the two-stage fusion of details and structure. This method utilizes DR projection images from two perspectives to achieve reconstruction while restoring high-quality details. Specifically, we will divide the reconstruction into two steps. Firstly, we utilize encoder-decoder architecture and cross feature map attention enhancement operator to achieve large-scale information recovery, thereby reconstructing device framework and other information; Secondly, introducing differential convolution to reconstruct small-scale detail information achieves the effect of detail enhancement. Finally, the two-step features are fused to output the 3D reconstructed volume. The algorithm also proposes an advanced loss variant that adds projection loss to common losses to improve reconstruction accuracy. To verify the effectiveness of the algorithm, we conducted validation on an open database named Chips_P_3D that we built ourselves. At the same time, to verify the generalization ability, we also conducted reconstruction experiments on fused datasets with complex information. The experiment shows that the reconstruction quality has been significantly improved.

Medical image segmentation, particularly for pathological diagnosis, faces challenges in leveraging patient clinical metadata that could enhance diagnostic accuracy. This study presents MMY-Net (Multimodal Y-shaped Network), a novel deep learning framework that effectively fuses patient metadata with pathological images for improved tumor segmentation performance. The proposed architecture incorporates a Text Processing Block (TPB) utilizing BERT for metadata feature extraction and a Text Encoding Block (TEB) for multi-scale fusion of textual and visual information. The network employs an Interlaced Sparse Self-Attention (ISSA) mechanism to capture both local and global dependencies while maintaining computational efficiency. Experiments were conducted on two open/public eyelid tumor datasets (Dataset 1: 112 WSIs for training/validation; Dataset 2: 107 WSIs as an independent test set) and the public Dataset 3 gland segmentation benchmark. For Dataset 1, 7989 H&E-stained patches (1024 & times; 1024, resized to 224 & times; 224) were extracted and split 7:2:1 (train:val:test); Dataset 2 was used exclusively for external validation. All images underwent Vahadane stain normalization. Training employed SGD (lr = 0.001), 1000 epochs, and a hybrid loss (cross-entropy + MS-SSIM + Lov & aacute;sz). Results show that integrating metadata-such as age and gender-significantly improves segmentation accuracy, even when metadata does not directly describe tumor characteristics. Ablation studies confirm the superiority of the proposed text feature extraction and fusion strategy. MMY-Net achieves state-of-the-art performance across all datasets, establishing a generalizable framework for multimodal medical image analysis.

Unsupervised visible-infrared person re-identification (VI-ReID) remains challenging because of the severe modality gap and the absence of identity annotations. Existing unsupervised VI-ReID methods often struggle to establish reliable cross-modal identity correspondence from noisy pseudo labels. They may also distort identity-related topology during modality alignment, which weakens discriminative representation learning. To enable reliable cross-modal identity discovery without manual annotations, we propose a cyclic pairwise identity learning framework for unsupervised VI-ReID. Our method first introduces a pedestrian pairwise relation constraint module. This module combines sample-level affinities with cluster-level identity cues to infer cross-modal pair-wise relations. The proposed framework further incorporates a cross-modal cyclic distribution alignment module. It performs invertible feature adaptation to reduce modality discrepancy while preserving identity-related topology. Extensive experiments validate the reliability and effectiveness of our proposed approach. The source code is available at https://github.com/ChenruiZhang267/CPIL.

Underwater image analysis is affected by light scattering, wavelength-dependent attenuation, low contrast, and suspended particles, which reduce the discriminative visual features. Current multi-scale Vision Transformers are not well-suited to these degradations because they cannot effectively fuse features across scales to achieve accurate classification. Although Vision Transformers (ViTs) can model long-range interactions, single-scale patch tokenization remains suboptimal for underwater images, where both fine-grained textures and global structures are important. This study proposes a Multi-Scale Vision Transformer (MS-ViT) with Cross-Scale Biased Attention Fusion (CSBAF) for underwater image classification. Before transformer encoding, the CSBAF introduces a learnable source–target scale-pair bias and an input-dependent scale-reliability gate. This differs from standard multi-scale fusion and cross-attention methods, which mainly concatenate features or exchange information between scale branches. The proposed design enables the model to emphasize reliable scales while suppressing degraded-scale responses. A hybrid dataset containing 14,000 images from the Roboflow Aquarium and RUIE datasets across five classes was used for evaluation. MS-ViT with CSBAF achieved 88.9% accuracy and an 88.8% F1-score, outperforming the CNN baseline by 7.6% and state-of-the-art transformer models, including UWFormer, DP-ViT, and CvT, by 2.3–4.2%. Ablation studies showed a 1.7% accuracy improvement over simple multi-scale concatenation, whereas cross-dataset testing achieved 84.4% accuracy, indicating reasonable cross-dataset robustness. These results demonstrate that explicit scale–aware fusion can improve transformer-based underwater visual understanding.

Wind power forecasting is critical for safe grid integration, yet wind-farm supervisory control and data acquisition time series are strongly non-stationary and spatially coupled, which often degrades accuracy and leads to unstable error growth at longer lead times. To address this challenge, we propose a physics-aware dynamic-graph mixture-of-experts (MoE) forecasting framework. The framework combines a hybrid CNN-Transformer temporal encoder to fuse local high-frequency disturbances with long-range and seasonal dependencies, a data-driven graph structure estimation module to learn time-varying turbine interactions, and a gated MoE that adaptively blends physics-informed and data-driven predictors under soft physical constraints to preserve plausibility. Experiments on real-world wind turbine operational data validated the method across ultra-short-term multi-step horizons from 1 h to 5 h and a representative medium-/long-term setting with a 10 day lead time at 15 min resolution. Compared with strong deep-learning baselines, the proposed approach reduces the mean squared error (MSE) by 25.7% and 21.2% on average over 1, 2, 3, and 5 h ahead forecasts, which improved robustness to error accumulation as the prediction horizon increases. Overall, coupling dynamic spatial learning with lightweight physics guidance can simultaneously improve accuracy and stability for dispatch-oriented wind power forecasting.

Frequency-division multiplexing (FDM) could effectively expand the high-frequency response capability of the phase-sensitive optical time-domain reflectometry (W-OTDR) system. However, the redistribution of the probe pulse energy into bilateral bands by Mach-Zehnder modulators significantly degrades the signal-to-noise ratio (SNR). A multiplex random coding (MRC) method is proposed for compensating the probe energy loss, which deploys distinct random coding pulses per channel to achieve SNR gain. Experimental results demonstrate that by using a five-channel sensing system, the SNR of the restored sinusoidal acoustic signals with a frequency of 15 kHz can reach 29.89 dB over a 10.438 km optical fiber. Compared with the conventional FDM W-OTDR system, the RMSE value of the demodulated signal at 45 km in a static environment decreased by 67.77 %. Furthermore, over a 48.536 km optical fiber, the SNR of the restored 1 kHz sinusoidal acoustic signal is improved by 9.37 dB, and the strain resolution is reduced from 120 p epsilon/ Hz to 31 p epsilon/ Hz. Compared with other FDM W-OTDR systems employing coding techniques, the scheme exhibits superior sensing performance. This method provides a potential solution for practical engineering applications of distributed acoustic sensing technology based on W-OTDR, particularly in detecting weak signals at high frequencies over a long distance.

Sparse-view CT mitigates clinical radiation risks by substantially reducing projection sampling, but this inevitably transforms image reconstruction into a highly ill-posed inverse problem. Although deep learning methods excel in image quality, they face critical bottlenecks: static supervised models suffer from overfitting and structural over-smoothing, whereas generative models struggle with high inference latency and physically inconsistent intensity drift. To address these challenges, this paper proposes a degradation-modulated Tri-Domain Synergistic Reconstruction Network (TD-SRN). The main contributions include: (1) Introducing a “many-to-one” dynamic manifold learning strategy. By constructing a continuous manifold of random input perturbations, it compels the network to extract robust anatomical structural invariances from diverse degraded states, alleviating the generalisation bottleneck on heterogeneous data. (2) Constructing a unified architecture that synergises the spatial, wavelet, and frequency domains. Specifically, a Radon-guided frequency module utilises projection geometry priors to explicitly compensate for missing spectral regions, mitigating streak artefacts and intensity drift from a physically informed perspective. Concurrently, a wavelet branch exploits multi-scale transform-domain sparsity to decouple high-frequency noise from fine anatomical details, effectively overcoming the traditional over-smoothing problem. Extensive experiments on the AAPM and DeepLesion datasets demonstrate that TD-SRN significantly outperforms other comparative methods in both quantitative metrics and visual quality under various sparse-view settings. This study provides a practical approach for sparse-view CT reconstruction, showing potential for real-time, high-fidelity clinical medical imaging.

Positron emission particle tracking (PEPT) enables non-intrusive Lagrangian trajectory measurements in opaque flow systems, yet accurate estimation of turbulent flow fields from PEPT data remains challenging due to uncertainty propagation inherent in numerical differentiation. Statistical characterization of a stationary tracer reveals that PEPT measurement uncertainties are well approximated by Gaussian distributions with pronounced spatial anisotropy. Based on this finding, a Gaussian-guided adaptive wavelet denoising framework is proposed to improve PEPT data quality. In this framework, the optimal denoising parameters are iteratively determined by verifying that the denoising residuals conform to Gaussian statistics, rather than relying on empirical selection. From the wavelet-denoised PEPT trajectories, instantaneous tracer kinematic quantities (location, velocity and acceleration), and 3D turbulent flow fields (mean velocity, mean acceleration and turbulent kinetic energy) are subsequently derived. Validation through a rotating motion benchmark confirms the effectiveness of the proposed denoising framework, achieving a reduction in the RMSE of location, velocity, and acceleration by 51%, 66.7%, and 97.2%, respectively. Further validation on a turbulent mixing flow confirms accurate estimation of 3D turbulent flow fields, showing excellent agreement between wavelet-denoised PEPT measurements and high-fidelity PIV measurements. The proposed denoising framework establishes PEPT as a reliable alternative tool for quantitative turbulence analysis, particularly in opaque fluids that optical techniques are infeasible.

Sparse-view Computed Tomography (CT), whilst an effective means of reducing radiation dose, is inherently a severely ill-posed inverse problem. Although deep learning-based methods have surpassed traditional algorithms in terms of image quality, they frequently suffer from over-smoothing or the risk of producing unreliable structures. More critically, existing diffusion models are predominantly confined to operations within the spatial domain, thereby neglecting crucial physical domain knowledge of CT imaging: the Fourier Slice Theorem (FST). This theorem dictates that sparse sampling corresponds directly to the absence of specific sectorial regions in the frequency domain. To address this theoretical gap, this paper proposes a physics-aware residual diffusion framework, termed FST-Diff. The primary contributions are as follows: (1) A Radon-Guided Frequency Modulation (RGFM) module is proposed. By utilising sinogram geometric embeddings to generate dynamic spectral gating masks, this module adaptively reduces spectral discrepancies associated with sparse sampling. (2) A Multi-scale Consistency Aggregation (MSCA) mechanism is constructed. Targeting the scale separation characteristics of global streaks and local noise, MSCA incorporates online physical re-projection during the training phase to enforce cross-scale geometric consistency. (3) A Proximal Consistency Equilibrium (PICE) inference strategy is designed. We re-model the reverse diffusion process as a proximal optimisation problem, employing the conjugate gradient method to solve for the optimal equilibrium point between the generative manifold and the measurement data. Experiments on the AAPM and DeepLesion datasets demonstrate that FST-Diff significantly outperforms other competing methods in both quantitative metrics and visual quality. Our dataset is available at https://github.com/hellorjb/FST-Diff.

The performance of deep learning (DL) models is highly dependent on the design of activation functions. However, traditional fixed-shape activation functions often suffer encounter issues such as gradient vanishing and local optimisation, particularly when dealing with complex tasks, thereby limiting their adaptability to diverse task requirements. Although trainable activation functions enhance the flexibility of DL models by incorporating learnable parameters, their optimisation process predominantly relies on gradient descent, which is prone to local optima. To address these limitations, this study proposes a data-driven, prior distributions-based optimisation framework for trainable activation functions. The proposed framework integrates a two-stage optimisation strategy, combining gradient descent and Bayesian inference, to significantly enhance neural network performance across multiple tasks. The core contributions of this paper are threefold: 1) The design of a generalised gated composite activation function that adaptively adjusts its shape according to task requirements by dynamically integrating multiple underlying activation functions. 2) The proposal of a two-stage optimisation framework that effectively alleviates the issue of local optima inherent in traditional optimisation methods. and 3) Comprehensive experimental validation on tasks such as image classification, regression, denoising, segmentation, and super-resolution, demonstrating that the proposed method delivers substantial performance gains across various tasks and datasets, surpassing existing classical activation functions and their variants. This study offers novel insights into the selection and optimisation of activation functions for DL models, holding significant academic and practical implications. Our dataset is available at https://github.com/hellorjb/GCAF.

High-density stacked semiconductor devices exhibit complex three-dimensional structures in industrial CT imaging, leading to insufficient defect localization accuracy when relying solely on single-slice 2D detection. To address this challenge, we propose a multi-class defect detection framework that integrates a feature-guided wavelet module (T-Wavelet) and a cross-dimensional attention fusion module (CDAF). The T-Wavelet module enhances subtle defect patterns through high-frequency amplification and low-frequency suppression while reducing background redundancy, providing structural priors for inter-slice feature integration. The CDAF module explicitly models spatial continuity among adjacent slices within a local depth range and maps 3D volumetric information into 2D channel representations, enabling unified 2D-3D feature modeling. Their combination forms the TW-CDAF module, which further strengthens local defect sensitivity and compensates for cross-layer structural discontinuities, thereby endowing a 2D detector with 3D structural perception. We construct a CT dataset of high-density stacked semiconductor devices covering multiple defect types. Experimental results demonstrate that the proposed framework achieves superior performance over mainstream detectors in key metrics such as mAP, validating its high accuracy and robustness under complex device structures.

High-resolution digital rock computed tomography (CT) images support pore-scale characterization of reservoir and gas-bearing porous media. However, micro-CT imaging is constrained by the trade-off between spatial resolution and field of view, and super-resolution (SR) reconstruction does not necessarily preserve pore structures even when pixel-level accuracy is high. Small gray-level errors near pore-matrix boundaries, narrow pore throats, and fracture-like structures can cause large changes in segmentation-derived connectivity and topology. To address this mismatch, we propose Pore-Aware Gaussian Splatting (PAGS), a continuous-coordinate SR framework that incorporates pore cues into 2D Gaussian Splatting reconstruction. A soft pore cue conditions Gaussian kernel assignment and query-level implicit decoding, allowing segmentation-sensitive regions to influence local Gaussian representation and final gray-value prediction. PAGS was evaluated at ×2 and ×4 on carbonate, sandstone, and coal datasets using image fidelity, threshold-dependent pore area recovery, region overlap, skeleton continuity, boundary-connected porosity, and Euler number deviation. PAGS achieved the highest peak signal-to-noise ratio (PSNR) in all six lithology scale combinations, the highest or joint-highest structural similarity index (SSIM) in all cases, and the highest Dice and clDice in every setting. Relative to the strongest evaluated baseline in each setting, PAGS reduced area porosity error by 8.6–59.6% and generally produced boundary-connected porosity responses closer to the high-resolution reference and smaller Euler number deviations. The topology-related advantage became less consistent under ×4 degradation, reflecting severe information loss. Overall, pore-aware continuous Gaussian reconstruction improves agreement between image-level SR accuracy and 2D segmentation-derived structural consistency, providing a more reliable image basis for downstream pore-scale characterization in digital rock workflows.

Acid fracturing technology is one of the most effective methods for resolving mineral plugging and for improving the pore structure of coal reservoirs. To investigate the characteristics of shallow and deep coal nanopore structures under the influence of acidic fracturing fluids, experiments using synchrotron radiation small-angle X-ray scattering were conducted on shallow and deep coal samples soaked in acidic fracturing fluids of different concentrations for varying durations. This quantitatively characterized the different nanoscale pore scattering intensity ratios (AI), fractal dimensions, and nanopore parameters. The research indicates that, under the influence of acidic fracturing fluids, the shallow coal nanopore structure tends to become more complex while that of deep coal becomes simpler. The impact of 20

Background and objective Sparse-view computed tomography (CT) speeds up scanning and reduces radiation exposure in medical diagnosis. However, when the projection views are severely under-sampled, deep learning-based reconstruction methods often suffer from over-smoothing of the reconstructed images due to the lack of high-frequency information. To address this issue, we introduce frequency domain information into the popular projection-image domain reconstruction, proposing a Tri-Domain sparse-view CT reconstruction model based on Sparse Transformer (TD-STrans). Methods TD-STrans integrates three essential modules: the projection recovery module completes the sparse-view projection, the Fourier domain filling module mitigates artifacts and over-smoothing by filling in missing high-frequency details; the image refinement module further enhances and preserves image details. Additionally, a multi-domain joint loss function is designed to simultaneously enhance the reconstruction quality in the projection domain, image domain, and frequency domain, thereby further improving the preservation of image details. Results The results of simulation experiments on the lymph node dataset and real experiments on the walnut dataset consistently demonstrate the effectiveness of TD-STrans in artifact removal, suppression of over-smoothing, and preservation of structural fidelity. Conclusion The reconstruction results of TD-STrans indicate that sparse transformer across multiple domains can alleviate over-smoothing and detail loss caused by reduced views, offering a novel solution for ultra-sparse-view CT imaging.

Micro-CT imaging is widely used in materials science, biomedical research, geological exploration, and industrial inspection due to its non-destructive capabilities. However, traditional systems face challenges due to limitations in spatial bandwidth product (SBP), making it difficult to achieve both a large field-of-view (FOV) and high-resolution imaging simultaneously. This paper presents a segmented light deflection method using cascaded optical fiber tapers integrated within a "fiber taper-image intensifier-camera" coupling architecture. This design reduces distortion caused by large taper ratios in single-stage systems. Additionally, a multi-view geometric registration calibration approach is developed based on CT projection geometric constraints to address mismatches between traditional 2D planar distortion correction methods and CT projection geometry. A cascaded fiber taper array (53 mm input - 34 mm intermediate - 16 mm output) was implemented in a 2 × 2 tiling configuration. The system achieves wide-FOV micro-CT imaging with a maximum imaging area of 106.4 × 106.4 mm2. Resolution testing confirmed a 3 μm resolution within a 4.1 × 4.1 mm2 FOV. System reconstruction efficacy was validated through rock core and printed circuit board (PCB) samples.

Multi-view X-ray imaging systems extend the imaging field of view by integrating multiple imaging modules. Nevertheless, inter-module performance variations in detector arrays, gain drift, and angle-dependent effects cause cross-view grayscale inconsistencies, which degrade overall imaging quality. While existing X-ray imaging approaches primarily rely on single-view flat-field correction or per-projection normalization, these strategies remain inadequate for achieving global consistency across views. To overcome this limitation, this study proposes a physics-constrained multi-view image correction (PCMIC) method. By constructing a globally shared attenuation field to model view-specific nonlinear responses and incorporating boundary anchoring together with monotonicity constraints, the method ensures strong physical fidelity. Experiments conducted on a self-developed large-field, high-resolution micro-CT system demonstrate that PCMIC effectively suppresses stitching artifacts and improves grayscale consistency, structural fidelity, and quantitative stability.

In cone-beam computed tomography (CT) systems, scatter signals can result in artifacts in images, thus adversely affecting image quality by reducing the contrast and signal-to-noise ratio. To accurately estimate and suppress scatter artifacts, a new scatter-correction method for cone-beam CT based on a beam-hole array (BHA) scatter-correction plate and the alternating direction method of multipliers (ADMM) are proposed. Based on the correction principle of the BHA scatter-correction plate, the plate is placed between the object under test and the detector during scanning. This setup captures projection data and computes scatter signals at the centers of the beam holes. Simultaneously, by incorporating the theory of compressed sensing, an L1-norm constrained model is established, and the ADMM is utilized to obtain a complete scatter-field distribution. Using an interpolation algorithm to scan angles reduces the time required for a full-angle reconstruction of the scatter field. Post-scatter removal and reconstruction yield scatter-corrected CT images. Data analysis of the reconstructed slices reveals the image quality, which indicates that the scatter-correction method based on the BHA plate effectively enhances the image contrast and reduces scatter artifacts. The improved scatter-signal reconstruction method further increases the accuracy of scatter-signal estimation, thus rendering scatter suppression more pronounced. Furthermore, using scanning-angle interpolation algorithms improves the efficiency of scatter-field reconstruction. The combined use of BHA and ADMM provides better scatter removal and further improvement in image quality.

Investigating the pore structure and understanding the relationship between pore characteristics and mechanical properties are crucial to research in the study of cement mortar. At present, the segmentation of large-scale concrete pores is mainly conducted using traditional algorithms or software, which are time-consuming and operate in a semi-automated manner. However, the application of these methods faces challenges when analyzing large-scale rock pores due to factors such as a lack of data, artifacts, and inconsistent contrast. In this study, six series of cement mortars were subjected to real-time CT scanning under uniaxial loading (RT-CT) to collect real-time three-dimensional data on the evolution of pore structures during loading. To address issues such as artifacts and inconsistent contrast, a new augmentation method was proposed to overcome artifacts and enhance contrast consistency. Finally, the augmented dataset was utilized for training, and the Fast R-CNN algorithm served as the framework for developing the pore recognition model. The results indicate that the improved algorithm demonstrates enhanced convergence and greater accuracy in pore segmentation. A mathematical model is developed to relate uniaxial compressive strength (UCS) to pore fractal dimension and porosity, based on pore segmentation analysis. The fractal dimensions evolution of each specimen is consistent with the progressive failure indicated by the strain-stress curve. Under uniaxial loading, specimens with a 4:1 cement–sand ratio exhibited peak strength. The incorporation of fractals improved particle contact, thereby facilitating the formation of the skeletal structure. These efforts contribute to improving the identification of the deformation of cement mortars.

Breast cancer is one of the most common malignant tumors among women, and early diagnosis can significantly mitigate its impact. Despite substantial advancements in breast cancer diagnosis using deep learning methods, many challenges persist. In clinical practice, transferring trained deep learning models to new, unlabeled patient samples is essential but challenging due to substantial variability among patient domains. Furthermore, existing domain adaptation models often neglect class-aware sub-domain gaps. Additionally, variations in image styles across domains further impede the accurate diagnostic in the target domain. To address these issues, we propose the category-weight instance fusion learning model for unsupervised domain adaptation in breast cancer diagnosis. This model employs a category-weighted contrast knowledge distillation module to align domains at the category level by selectively clustering similar samples and segregating dissimilar ones. Simultaneously, the meticulously designed instance-aware feature mixing module merges image styles across domains through a domain feature mixing algorithm, significantly enhancing breast cancer domain adaptation capability. Results on BreakHis and ICIAR-2018 datasets demonstrate that our model outperforms other state-of-the-art domain adaptation algorithms in diagnostic accuracy, proving the transferability and robustness of our model across diverse clinical patients.

Computed tomography(CT)technology has demonstrated significant application value in industrial inspection owing to its non-destructive testing capabilities,high resolution,and visualization features.However,in certain industrial inspection scenarios,extremely limited scanning conditions pose substantial challenges for projection data acquisition,restricting the application of traditional reconstruction methods.To address this challenge,this study proposes an orthogonal dual-view 3D reconstruction network tailored for rapid CT imaging.The proposed method employs an encoder-decoder architecture,utilizing 2D convolutions instead of 3D convolutions to infer the depth dimension of CT volumes through feature channels,thereby enhancing model inference speed.Additionally,gradient information and gradient loss are introduced to strengthen the edge recovery capability of the network.The method is validated on walnut and Fuze datasets.Experimental results showed that reconstructing a volume with a resolution of 128 required only 0.19 s,and the structural similarity of the reconstructed images was higher than 0.98.This approach demonstrates effective capability in inferring 3D CT volumes from dual-view 2D projections,revealing its future potential in rapid CT imaging.