A deep learning framework integrating SMOTE algorithm and GC e-nose for tracing the geographical origins of food: taking Astragali Radix as an example.

Glaucoma, the second largest cause of irreversible blindness worldwide, causes significant damage to the optic nerve. Early diagnosis of glaucoma is crucial since without it, there will be continuous deterioration of vision. Manual detection of glaucoma based on fundus images is a tedious and potentially inaccurate process, making it important to develop computer-aided glaucoma detection systems. This paper introduces a hybrid approach to glaucoma detection that involves the use of Local Binary Pattern (LBP) and Pivot Distribution Count (PDC) feature extraction methods for the analysis of retinal fundus images. LBP is a technique that involves the extraction of texture-based features, whereas PDC is a feature extraction algorithm that involves the extraction of white-pixel intensity and fractal dimension from retinal fundus images. In this study, features were extracted and classified using different machine learning algorithms such as SVM, Decision Trees (DT), Random Forest (RF), KNN, Adaboost, Gradient Boosting, XGboost, Light Gradient Boosting Machine, and CatBoost. Moreover, grid search cross validation, randomized search cross validation, genetic algorithm, and Bayesian optimization were also applied to optimize the performance of LightGBM and Catboost classifiers. The proposed Hybrid LBP-PDC method provided maximum classification accuracy of 97.55% using the CatBoost and LightGBM classifiers. Moreover, we have developed a robust and efficient methodology for the automatic glaucoma screening using the retinal fundus image analysis.

The biosensors based on metasurfaces have attracted significant attention for biological detection in the terahertz (THz) band due to the characteristics of rapid, label-free, and nondestructive. Benefiting from the low-loss characteristics of the electromagnetically induced transparency (EIT), we propose two single-band EIT meta-biosensors and a dual-band EIT meta-biosensor through various combinations of three designed T-shaped resonators that act as bright modes. By utilizing the excellent biocompatibility of gold nanoparticles (AuNPs), the proposed meta-biosensors modified with carcinoembryonic antigen (CEA) antibody-conjugated AuNPs achieve the specific detection of CEA within mixed tumor markers in the THz band, which indicates that the frequency shifts of the transparent peaks increase only with increasing CEA concentration, as the CEA antibody-conjugated AuNPs specifically capture the target antigens. Furthermore, we employ a mutual information feature extraction algorithm integrated with a support vector machine (SVM) to verify the discrimination of transmission spectra samples with different CEA concentrations, and the classification results show high intraclass consistency and clear interclass distinction. This work paves the way for the development of novel meta-biosensors with high sensitivity and specific target antigen recognition, which would provide technical support for cancer screening and disease diagnosis.

Diagnostic accuracy in medical imaging depends on high-quality multimodal image fusion (MMIF). It improves image quality by combining data from diverse imaging modalities. However, many existing fusion approaches fail to capture distinct information from diverse modalities, resulting in an incomplete fused image and potentially a misleading, incorrect diagnosis. To address these issues, we propose a new approach to the MMIF framework that incorporates a neutrosophic fuzzy set (NFS) with a feature extraction and arithmetic optimization algorithm (AOA), which are designed to eliminate the uncertainty and indeterminacy present in medical modalities. In the first phase, we utilized a Gaussian filter to decompose the source modalities into distinct layers, including the base and detail. In the second phase, the detail layers were fused with the optimal weights, generated by the AOA fusion strategy, which preserved the significant edge features and retained the comprehensive information. In the third phase, the base layers were converted into neutrosophic fuzzy images (NFI) using a neutrosophic fuzzy set. Thereafter, the α-mean and β-enhancement operations were employed to enhance the quality of the NFI images. Then, the Tamura feature extractions and contrast visibility enhancements were implemented for the extraction of significant features from the modalities of the base layers. In the last phase, the final resultant image was obtained by integrating the fused detail and base components. Experimental outcomes reveal that the presented method outperforms other fusion approaches. The qualitative and objective evaluation demonstrates that the proposed approach yields a composite image with an excellent visual appearance and contrast without artifacts. Moreover, this developed method consistently yields better objective values while maintaining reasonable runtime analysis, thereby balancing efficiency as well as fusion performance.

ABSTRACT This study proposes a deep learning method based on an improved PointNet++ network for point-cloud segmentation and volume estimation of dense medium piles in coal preparation plants. By optimizing the original PointNet++ model, two main improvements are introduced: an enhanced ball neighborhood feature extraction algorithm and a simplified network architecture. These improvements significantly enhance segmentation accuracy under non-uniform point density and complex local morphologies, while also increasing computational efficiency. Experimental results demonstrate that, compared with traditional methods such as plane clipping, RANSAC plane fitting, and progressive morphological filtering, the improved model achieves higher segmentation accuracy, particularly at pile boundaries and slope regions. The mean mIoU increases from 88.94% to 90.27%, with improvements mainly observed in critical boundary regions and areas affected by foreign objects, thereby reducing volume calculation errors and improving system reliability. In addition, the processing time is reduced by 50%. For volume estimation, a slicing method based on the segmented point cloud is used. The results show that, with a slice thickness of 1.375 mm, the volume calculation error is minimized, and in practical applications, the volume estimation error is consistently kept below 10%. This study provides a theoretical foundation and practical approach for medium replenishment and intelligent grasping systems in coal preparation plants, with broad industrial application prospects.

HighlightsWhat are the main findings?Preprocessing attenuated color features in limited-band field in situ spectra.CARS-SVR effectively avoids high-value underestimation from data imbalance.What are the implications of the main findings?RF with physical indices enables low-cost, effective low-content SOM estimation.A synergistic approach serves as a reference for forest soil SOM estimation.Rapid and effective estimation of soil organic matter (SOM) is crucial for the scientific management of Moso bamboo forests. This study investigated Moso bamboo forest soils in Yongan City, Fujian Province, and systematically evaluated the synergistic adaptation strategies coupling spectral preprocessing methods, feature extraction strategies, and machine learning models based on visible and shortwave near-infrared (Vis-NIR) spectroscopy. The results indicated that: (1) Conventional preprocessing algorithms attenuated the SOM spectral feature signals dominated by soil color within the limited wavelength range of field in situ spectral data, resulting in a general decline in the accuracy of the estimation models. (2) Feature extraction and modeling algorithms exhibited distinct adaptability across different content intervals. Within the low-content interval (<15 g/kg), simple physical indices combined with random forest (RF) achieved effective estimation at a lower computational cost (RPD = 2.18). Within the high-content interval (>25 g/kg), the synergistic strategy of the CARS algorithm combined with support vector regression (SVR) yielded the optimal estimation performance (R2 = 0.83, RPD = 2.45) and effectively mitigated the underestimation of high values caused by data imbalance. In conclusion, this study proposed a feature–model synergistic estimation approach, validating its feasibility for estimating SOM in Moso bamboo forests under the specific constraints of the current study area, thereby serving as a valuable reference for forest soil SOM monitoring in specific regions.

Feature extraction and state recognition algorithm for DC servo motors based on MPE-VMD and CSSVM

Deep Learning-Based Positive Region Segmentation and Spatial Registration of Virtual Multiplex Immunohistochemical Whole-Slide Images.

Image Super-Resolution (ISR) is employed to generate high-resolution images from low-resolution inputs. However, most current techniques for ISR encounter important challenges such as: (i) the assumption of sufficient training data availability, and (ii) the presumption that target image regions are complete without missing data. To address these practically important challenges, this study applies a lightweight approach termed Fuzzy Rough Feature Selection-based ANFIS Interpolation for ISR, especially on Martian imagery. Feature extraction algorithms are first applied to capture potentially significant features, and population-based search mechanisms are then utilised to perform effective feature selection (via extending the popular fuzzy-rough feature selection mechanism). The selected feature set is subsequently fed into an ANFIS interpolation model to perform the ISR task. Particularly, to handle the issue of sparse and incomplete data in dealing with Mars images, two adjacent ANFIS models are trained on nearby regions with sufficient data, positioning the model for the sparse region in between. Experimental studies conducted on Martian image datasets under both sufficient and sparse data conditions validate the effectiveness of the proposed approach, in overcoming the specific challenges faced by the task of ISR in extraterrestrial imaging scenarios.

Pegmatites are a major source of lithium (Li), a critical metal enabling the transition from a fossil fuel-based energy sector to a sustainable, renewable energy future. The Greenbushes Li deposit in Western Australia is the world’s largest single source of Li, and many other geological environments across Australia are highly prospective for pegmatite-hosted Li deposits. The recent Li rush has highlighted that downstream industries underestimate the complexity of pegmatite-hosted Li deposits, particularly regarding the wide range of Li-bearing minerals present in a single pegmatite. This paper presents an overview of Li-pegmatite mineral assemblages from various regions across Australia, including the southwestern part of Western Australia (WA), hosting Greenbushes, the Eastern Yilgarn Craton (Londonderry), the Pilbara Craton (Pilgangoora, King Col), Queensland’s Georgetown Inlier (Buchanan’s Creek) and the Dorchap Pegmatite Swarm in Victoria. Mineral assemblages from said localities are examined using publicly available hyperspectral drill core analyses (HyLogger3™). These findings are compared with those of Greenbushes drill core C3DD0024, supported by whole-rock geochemistry, Fourier transform infrared (FTIR) spectroscopy and X-ray diffractometry (XRD). Reflectance spectral characteristics of Li-bearing, gangue and alteration minerals are evaluated across the visible near-infrared (VNIR), shortwave infrared (SWIR) and thermal infrared (TIR) wavelength ranges to identify suitable unmixing and feature extraction algorithms for inferring relative mineral abundances and mineral chemistry. Newly developed algorithms—most notably the spodumene abundance index, show a significant correlation with XRD validation and whole-rock geochemistry, demonstrating their effectiveness for mineral detection. Furthermore, the distribution of other Li-bearing minerals, such as petalite and eucryptite, was mapped across the study sites using reference library spectra. The applied hyperspectral mineral abundance and composition indices provide key insights into mineral assemblages and their variations by (1) tracing variations of mineral assemblages within pegmatites and country rocks at centimetre scale, (2) showing that spodumene is predominantly associated with quartz ± plagioclase (without K-feldspar), whereas petalite is mostly associated with quartz and K-feldspar in the absence of plagioclase, (3) confirming that most white mica in pegmatites exhibits an Al-rich composition, (4) demonstrating that phengitic alteration halos around pegmatites are narrow and (5) indicating that chlorite ± biotite ± epidote alteration is common in mafic–ultramafic and metasedimentary wallrocks. The algorithms and workflows provided in this study offer significant potential for expanding applications to large-scale hyperspectral drill core datasets, furthering the understanding of pegmatite genesis, supporting Li-exploration and guiding metallurgical processes by characterising Li-host minerals.

Background:Analyzing neural data such as electroencephalography (EEG) data often involves dealing with high-dimensional datasets, where not all channels provide equally meaningful information. Selecting the most relevant channels is crucial for improving computational efficiency and ensuring robust insights into neural dynamics.Method:This study introduces the Importance of Channels based on Effective Connectivity (ICEC) criterion for quantifying effective connectivity (EC) in each channel. EC refers to the causal influence one neural region exerts over another, providing insights into the directional flow of information. Using this criterion, we propose an unsupervised channel selection method that accounts for the intensity of interactions among channels.Results:To evaluate the proposed channel selection method, we applied it to three well-known EEG datasets across four categories. The assessment involved calculating the ICEC criterion using five EC metrics: partial directed coherence (PDC), generalized PDC, renormalized PDC, directed transfer function (DTF), and direct DTF. To focus on the effect of channel selection, we employed the common spatial pattern algorithm for feature extraction and a support vector machine for classification across all participants. We compared our results against other CSP-based methods. The evaluation included comparing participant-specific accuracies with and without the proposed method across five EC metrics.Conclusion:The results showed consistent improvements and a significant reduction in the number of electrodes selected for all participants. Compared to state-of-the-art methods, our approach achieved the highest accuracies: 82% (13 out of 22 channels), 86.01% (29 out of 59 channels), and 87.56% (48 out of 118 channels) across all three datasets.

UAV-based deep learning for biodiversity monitoring: Advances, applications, and future directions

Research on feature extraction algorithms for high-dimensional substation data based on heterogeneous weighted bayesian optimization

Camel milk powder, as a premium product, is often diluted with inferior substitutes, specifically through the adulteration with ordinary milk powders. Due to the high similarity in composition among different types of milk powders, detecting adulterated camel milk powder presents significant challenges. This study establishes a discrimination analysis model for pure camel milk powder and mixtures containing camel, goat, and cow milk powders based on hyperspectral technology combined with an improved Black-winged Kite algorithm. Firstly, spectral data preprocessing was performed using single and combined spectral preprocessing methods; ultimately, the S-G-SNV method was selected. Secondly, a feature extraction algorithm based on one-dimensional dilated convolutional neural networks (1D Dilated CNNs) was proposed to facilitate feature fusion and enhance model accuracy. On this basis, a multi-class qualitative analysis model utilizing an improved black-winged kite algorithm (IBKA) applied to SVC was constructed. The experiment compared five qualitative analysis models; the final S-G-SNV-1D Dilated CNNs-IBKA-SVC model achieved an accuracy of 0.9493, an F1 score of 0.9478, and Cohen's kappa value of 0.947 on the test set—demonstrating superior performance over other methods. This model provides a novel and effective solution for the rapid non-destructive identification and analysis of milk powder from different sources. • .An improved BKA algorithm has been proposed. • A 1D Dilated CNNs algorithm is proposed for spectral feature extraction. • Comparative Analysis of different Preprocessing Methods (Single and Combined). • SHAP provides interpretable explanations of feature wavelength contribution rates. • Non-destructive high-precision identification of adulterated camel milk powder.

With the development of English education and teaching, the use of digital technology in oral classroom teaching has been emphasized. In this paper, digitalization empowers English speaking training and explores the innovative mode of English speaking training. Based on the application of teaching platform and differentiated teaching needs therein, the English speaking learning system based on personalized needs is proposed, and a personalized resource recommendation model is constructed through the combination of student interest feature extraction and collaborative filtering recommendation algorithm. Then it carries out the teaching practice of students in four business English major classes in a school. From the experiment, it is evident that this system offers a higher accuracy and faster resource recommendation to the experimental group compared to the other system with an accuracy of more than 90% and a recommendation time decrease of 79.51% and 72.50%, respectively. The improvement of students' oral performance and oral expression time in the experimental group were 9.82%~15.09% and 150.21%~176.05%, respectively, and the class with this paper's system had the best performance among the three classes. The English speaking training model based on digitization can promote students' willingness to express English and the length of expression, and enhance students' English speaking learning effect.

Hyperspectral imaging, as a novel non-contact, non-ionizing radiation imaging technique, can acquire abundant spectral–spatial information without damaging tissue. Addressing the critical challenge of precisely delineating boundaries during brain glioma surgery, this study proposes a new hyperspectral-assisted diagnostic and classification method for brain gliomas based on cross-channel fusion feature extraction. The method acquires tissue spectral–spatial information using hyperspectral imaging, employs a spatial–spectral fusion channel feature extraction algorithm, and designs a multi-branch feature extraction network to achieve precise discrimination between tumor and normal tissue. The network uses a multi-scale feature extraction architecture: the main branch extracts deep semantic features via composite convolutional units, while a side branch preserves the original spectral–spatial information and integrates features through residual connections, thereby substantially enhancing the model’s representational capacity and robustness. The advantages of this method lie in its balance of local detail and global semantics; by complementing information across channels, it effectively strengthens feature discriminability, while also offering good generalization and interpretability. Experimental results show a classification accuracy of 97% on the validation set, a maximum accuracy of 96% on an external test set, and an average accuracy of 90%. ROC curve analysis indicates stable model performance, with AUC values up to 0.99 and an average of 0.95, significantly outperforming traditional methods. Further analysis confirms that the model effectively captures discriminative features in key spectral bands (e.g., 590–610 nm and 650–700 nm), which are highly related to hemoglobin absorption characteristics, thus providing a clear biophysical explanation for the classification results. This study not only validates the effectiveness of hyperspectral imaging combined with deep learning for brain glioma recognition, but the proposed multi-branch architecture also offers a new technical pathway for medical hyperspectral image analysis.

Adolescent mental health has emerged as a critical public health challenge, with traditional screening methods often hindered by time lags and limited reach. As social media becomes a primary channel for emotional expression among youth, automated sentiment analysis offers a promising pathway for real-time monitoring. This study constructs an automated recognition system using a large-scale corpus of 52,573 labeled social media entries across seven psychological dimensions, including depression, anxiety, suicidal ideation, and stress. By employing the TF-IDF algorithm for multi-dimensional feature extraction and a Logistic Regression model for multi-class classification, the proposed scheme achieves an overall recognition accuracy of 74.8%. Experimental results reveal significant linguistic patterns across mental states: the "normal" category exhibited the highest discriminability (F1-score = 0.896), while the "stress" category proved the most challenging to identify (F1-score = 0.553) due to its semantic overlap with daily emotional fluctuations. Feature analysis further confirms that specific "psychological fingerprints"—such as the high frequency of first-person pronouns in the depression group and uncertainty-related queries in the anxiety group—can serve as reliable predictors. This research validates the feasibility of large-scale, non-invasive psychological screening and provides a data-driven framework for early campus crisis intervention and precise psychological support.

Commutation failure poses a significant operational risk on HVDC transmission systems, with its prediction facing challenges from complex underlying mechanisms and stringent real-time requirements. Data-driven techniques show promise, but practical implementation of fully data-driven solutions remains unresolved. This paper introduces a novel fully data-driven fast prediction framework for commutation failures, featuring three principal innovations: (i) A second-order determinant-based feature extraction algorithm that compresses data dimensionality while preserving critical disturbance characteristics; (ii) A cost-sensitive learning technique integrated with sample augmentation strategies to address sample bias problem; (iii) A classification performance evaluation protocol tailored for engineering applications. Experimental validation on the modified CIGRE benchmark system demonstrates 1.25 and 1% false alarm rate and miss rate respectively, with response time reduced to within 2 ms (at 2.5 kHz sampling rate). The proposed methodology significantly reduces dependency on extensive training datasets, offering a viable purely data-driven solution for real-time commutation failure pre-judgement.

As the core infrastructure of national comprehensive transportation system, the safe operation of railway lines is crucial. Quadruped robots have become ideal carriers for railway autonomous inspection due to their excellent terrain adaptability. High-precision environmental perception is the core prerequisite for their autonomous operation. Aiming at the defects of traditional PV-RCNN algorithm in point cloud feature extraction for railway scenarios, this paper proposes an improved TS-PV-RCNN algorithm for 3D object detection using LiDAR point cloud. By introducing transform-equivariant strategy, dual attention mechanism, and Transformer feature extraction module, the feature extraction effect is optimized. Experiments on KITTI dataset and railway field dataset show that compared with the benchmark PV-RCNN algorithm, the average precision (AP) of railway equipment/vehicles, pedestrians, and foreign objects is improved by 1.35%, 16.83%, and 9.24% respectively under medium difficulty. The proposed algorithm provides a feasible technical scheme for autonomous inspection of railway lines.

Advances in medical imaging have led to massive archives, yet navigating these datasets remains challenging due to the limitations of traditional text-based search engines. While content-based medical image retrieval (CBMIR) offers a visual feature-based solution to enhance clinical workflows and research, its operational integration into picture archiving and communication systems (PACS) remains a significant bottleneck. Despite the progress in deep learning for feature extraction, CBMIR tools are rarely integrated and effectively implemented within existing radiology information systems due to complex protocol barriers. To address these challenges, this study develops a CBMIR system meticulously designed to cater to 7 distinct types of brain tumors as seen in brain magnetic resonance images. Our system is tailored to assist radiologists and health care professionals in efficiently retrieving pertinent historical medical images, thereby providing quantitative decision support for radiologists and facilitating evidence-based case comparison, with the potential to improve retrieval efficiency and clinical workflow, rather than directly enhancing diagnostic accuracy. The dataset used in this study was collected from a single medical center and is not publicly available. The core innovation is a state-of-the-art deep learning-based feature extraction algorithm specifically engineered for the CBMIR system. We use GoogLeNet as the primary architecture, incorporating generalized mean pooling to capture nuanced local features and an embedding layer for dimension reduction. Crucially, we address the integration gap by harmonizing 2 open-source projects to successfully embed the CBMIR system into a functional PACS environment via standard protocols. The image dataset contains 658 participants with 15,873 images collected from 2000 to 2017. The empirical findings of our research demonstrate the performance and robustness of the proposed CBMIR system. Our system achieves a remarkable mean average precision score of 89.16% and an equally impressive Precision@10 score of 94.08%. These metrics affirm the system's efficacy in retrieving relevant medical images. Furthermore, we successfully integrate the CBMIR system into a PACS by successfully harmonizing 2 open-source projects. This study presents the design and implementation of a PACS-integrated CBMIR system for brain magnetic resonance imaging, and the experimental results demonstrate that the system can achieve efficient and accurate image retrieval within a clinical workflow.