Dual-Branch Attention-based Frequency Domain Network for Cross-subject SSVEP-BCIs.

Objective. To enhance the decoding accuracy and information transfer rate of steady-state visual evoked potential-based brain-computer interface (SSVEP-BCI) systems and to reduce inter-subject variability for broader SSVEP-BCI applications, a dual-channel TRCA-net (DC-TRCA-net) method is proposed, based on cross-subject positive transfer. The proposed method incorporates an innovative Transfer-Accuracy-based Subject Selection (T-ASS) strategy and a deep learning network integrated with the SSVEP Domain Adaptation Network (SSVEP-DAN) to enhance SSVEP-BCI decoding performance. The T-ASS strategy constructs contribution scores by computing each subject's self-accuracy and transfer accuracy, and enables effective source subject selection while mitigating negative transfer risks. DC-TRCA-net is further developed to improve model generalization through cross-subject data augmentation. The effectiveness of the proposed method is validated on two large-scale public benchmark datasets. Experimental results demonstrate that DC-TRCA-net outperforms existing networks across both datasets, with particularly substantial performance gains observed in complex experimental scenarios.

Enhancing detection of SSVEPs using discriminant compacted network

High-frequency steady-state visual evoked potential-based brain-computer interface (SSVEP-BCI) systems offer improved user comfort but suffer from reduced performance compared to their low-frequency counterparts, limiting their practical application. To address this issue, we propose a transfer learning-based method that leverages low-frequency SSVEP data to enhance high-frequency SSVEP performance. A filtering mechanism is designed to extract informative components from low-frequency signals, and the least squares algorithm is employed to generate high-quality synthetic high-frequency data. Experiments conducted on two public datasets using TDCA, eTRCA, and advanced TRCA-based algorithms demonstrate significant performance improvements. Our approach requires only two calibration trials, achieving 9.03% and 14.49% accuracy increases for eTRCA and TDCA in Dataset 1, and 13.91% and 14.53% improvements in Dataset 2, all within 1.5s. Moreover, our approach effectively addresses the issue of single calibration data for high-frequency SSVEP-BCI systems. These results support the feasibility of fast calibration and improved performance in real-world high-frequency BCI applications.

Learning informative and discriminative semantic features for robust facial expression recognition

Accurate histological subtype classification between adenocarcinoma (ADC) and squamous cell carcinoma (SCC) using computed tomography (CT) images is of great importance to assist clinicians in determining treatment and therapy plans for non-small cell lung cancer (NSCLC) patients. Although current deep learning approaches have achieved promising progress in this field, they are often difficult to capture efficient tumor representations due to inadequate training data, and in consequence show limited performance. In this study, we propose a novel and effective reconstruction-assisted feature encoding network (RAFENet) for histological subtype classification by leveraging an auxiliary image reconstruction task to enable extra guidance and regularization for enhanced tumor feature representations. Different from existing reconstruction-assisted methods that directly use generalizable features obtained from shared encoder for primary task, a dedicated task-aware encoding module is utilized in RAFENet to perform refinement of generalizable features. Specifically, a cascade of cross-level non-local blocks are introduced to progressively refine generalizable features at different levels with the aid of lower-level task-specific information, which can successfully learn multi-level task-specific features tailored to histological subtype classification. Moreover, in addition to widely adopted pixel-wise reconstruction loss, we introduce a powerful semantic consistency loss function to explicitly supervise the training of RAFENet, which combines both feature consistency loss and prediction consistency loss to ensure semantic invariance during image reconstruction. Extensive experimental results show that RAFENet effectively addresses the difficult issues that cannot be resolved by existing reconstruction-based methods and consistently outperforms other state-of-the-art methods on both public and in-house NSCLC datasets. Supplementary material is available at https://github.com/lhch1994/Rafenet_sup_material.

Motor imagery (MI)-based brain-computer interface (BCI) systems are being increasingly employed to provide alternative means of communication and control for people suffering from neuro-motor impairments, with a special effort to bring these systems out of the controlled lab environments. Hence, accurately classifying MI from brain signals, e.g., from electroencephalography (EEG), is essential to obtain reliable BCI systems. However, MI classification is still a challenging task, because the signals are characterized by poor SNR, high intra-subject and cross-subject variability. Deep learning approaches have started to emerge as valid alternatives to standard machine learning techniques, e.g., filter bank common spatial pattern (FBCSP), to extract subject-independent features and to increase the cross-subject classification performance of MI BCI systems. In this paper, we first present a review of the most recent studies using deep learning for MI classification, with particular attention to their cross-subject performance. Second, we propose DynamicNet, a Python-based tool for quick and flexible implementations of deep learning models based on convolutional neural networks. We show-case the potentiality of DynamicNet by implementing EEGNet, a well-established architecture for effective EEG classification. Finally, we compare its performance with FBCSP in a 4-class MI classification over public datasets. To explore its cross-subject classification ability, we applied three different cross-validation schemes. From our results, we demonstrate that DynamicNet-implemented EEGNet outperforms FBCSP by about 25%, with a statistically significant difference when cross-subject validation schemes are applied.

Proton nuclear magnetic resonance (NMR) spectroscopy provides a powerful tool for chemical profiling, also known as spectral fingerprinting, because of its inherent reproducibility. NMR is now increasing in use for authentication of complex materials. Typically, the absorbance spectrum is used that is obtained as the phase-corrected real component of the Fourier transform (FT) of the free induction decay (FID). However, the practice discards half the information that is available in the dispersion spectrum obtained as the imaginary component from the FT. For qualitative analysis or quantitative analysis of small sets of absorbance peaks, the symmetric and sharp peaks of the real spectra work well. However, for pattern recognition of entire spectra, trading peak resolution for peak reproducibility is beneficial. The absolute value of the complex spectrum gives the length or magnitude of magnetization vector in the complex plane; therefore, the magnitude relates directly to the signal (i.e., induced magnetization). The magnitude spectrum is obtained as the absolute value from the real and imaginary spectral components after the FT of the FID. By breaking with tradition and using the magnitude spectrum the reproducibility of the spectra and consequent recognition rates can be improved. This study used a 500-MHz 1H NMR instrument to obtain spectra from 4 diverse datasets; 12 tea extracts, 8 liquor samples, 9 hops extracts, and 25 Cannabis extracts. Six classifiers were statistically evaluated using 100 bootstrapped Latin partitions. The classifiers were a fuzzy rule-building expert system (FuRES) tree, support vector machine trees (SVMTreeG and SVMTreeH), a regularized linear discriminant analysis (LDA), super partial least squares discriminant analysis (sPLS-DA), and a one against all support vector machine (SVM). All classifiers gave better or equivalent results for the magnitude spectral representation than for the real spectra, except for one case of the 24 evaluations. In addition, the enhanced reproducibility of the absolute value spectra is demonstrated by comparisons of the pooled within sample standard deviations. For pattern recognition of NMR spectra, the magnitude spectrum is advocated.

One of the historical challenges in the field of passive vibration damping technology is the modeling of experimental complex modulus data as a function of temperature and frequency. Several models exist for the temperature shift function (TSF); i.e., Arrhenius, WLF, exponential, etc. Several models of complex modulus as a function of reduced frequency also exist. All existing TSF and CM models fail to represent at least some sets of data with desired accuracy and efficiency. Consider a linear, constant coefficient, stable system, and its frequency response function. It is well known that if the real component of the complex‐valued frequency response function is given over the infinite frequency range, then the imaginary component may be obtained. The complex modulus of vibration damping materials is such a system. Extensive work with fractional calculus based models for complex modulus has established their viability and potential attractiveness. A ratio of factored polynomials of one‐half order is proposed to model the complex modulus. This CM model is attractive from a number of viewpoints: The proper interrelationship of the real and imaginary components is guaranteed; an adequately large number of terms may be used in order to accurately model the complex modulus; an expression may be developed for the real component that lends itself to fitting data by collocating through a number of points; closed‐form expressions may be developed for compliance, relaxation modulus, and creep compliance which also lend themselves to collection fitting of experimental data, etc. With modern computational power, this model becomes both accurate and efficient. Previous work has established the slope of the TSF as the characteristic which causes complex modulus data to be properly shifted; therefore, a new approach to modeling the TSF is proposed. The new model is based on determining values of slopes at equally spaced temperatures, fitting a cubic spline through these points (i.e., knots), storing the coefficients, and integrating the cubic spline analytically. The concept of reduced temperature is introduced, used as a convenience for the present effort and proposed as an additional method of presenting data in a form useful to the damping industry. The core of the revolutionary concept is using the simultaneous modeling of both real and imaginary components as the criteria to enable the set of data to establish its TSF. Previous techniques have used real modulus, imaginary modulus, and loss factor as a function or reduced frequency, sometimes in a least‐squares sense, and sometimes visually, as the criteria. The above CM and TSF models are essential to the iteration strategy required to determine parameter values for both models. The iteration scheme is conceptually straight‐forward. Approximations to the CM and to the TSF are obtained. For each TSF knot, the reduced temperature is used to determine the associated reduced frequency, the current loss factor curve is compared to the corresponding experimental value and the value of the slope adjusted accordingly, the real component collocated for the updated TSF, etc. Examples are given and discussed.

COVID-19, an infectious coronavirus disease, caused a pandemic with countless deaths. From the outset, clinical institutes have explored computed tomography as an effective and complementary screening tool alongside the reverse transcriptase-polymerase chain reaction. Deep learning techniques have shown promising results in similar medical tasks and, hence, may provide solutions to COVID-19 based on medical images of patients. We aim to contribute to the research in this field by: (i) Comparing different architectures on a public and extended reference dataset to find the most suitable; (ii) Proposing a patient-oriented investigation of the best performing networks; and (iii) Evaluating their robustness in a real-world scenario, represented by cross-dataset experiments. We exploited ten well-known convolutional neural networks on two public datasets. The results show that, on the reference dataset, the most suitable architecture is VGG19, which (i) Achieved 98.87% accuracy in the network comparison; (ii) Obtained 95.91% accuracy on the patient status classification, even though it misclassifies some patients that other networks classify correctly; and (iii) The cross-dataset experiments exhibit the limitations of deep learning approaches in a real-world scenario with 70.15% accuracy, which need further investigation to improve the robustness. Thus, VGG19 architecture showed promising performance in the classification of COVID-19 cases. Nonetheless, this architecture enables extensive improvements based on its modification, or even with preprocessing step in addition to it. Finally, the cross-dataset experiments exposed the critical weakness of classifying images from heterogeneous data sources, compatible with a real-world scenario.

Finer-grained local features play a supplementary role in the description of pedestrian global features, and the combination of them has been an essential solution to improve discriminative performances in person re-identification (PReID) tasks. The existing part-based methods mostly extract representational semantic parts according to human visual habits or some prior knowledge and focus on spatial partition strategies but ignore the significant influence of channel information on PReID task. So, we proposed an end-to-end multi-branch network architecture (MCSN) jointing multi-level global fusion features, channel features and spatial features in this paper to better learn more diverse and discriminative pedestrian features. It is worth noting that the effect of multi-level fusion features on the performance of the model is taken into account when extracting global features. In addition, to enhance the stability of model training and the generalization ability of the model, the BNNeck and the joint loss function strategy are applied to all vector representation branches. Extensive comparative evaluations are conducted on three mainstream image-based evaluation protocols, including Market-1501, DukeMTMC-ReID and MSMT17, to validate the advantages of our proposed model, which outperforms previous state-of-the-art in ReID tasks.

In cryoEM 3D reconstruction, noise is spread over the entire 3D space, while the signal is present only within the molecular mask. The signal-to-noise ratio (SNR) within the molecular mask is the SNR of interest in data analysis. Any method that does not apply the molecular mask to noise underestimates the SNR. For instance, Fourier Shell Correlation (FSC) curves recalculated between two half-maps after the data are deposited in the PDB do not apply the molecular mask to noise, consequently providing a misleading impression of resolution. These approaches, relying on calculations without mask constraints on noise, try to avoid introducing biases within the mask area that could potentially arise at low resolution. Here, we present a method of FSC calculations in which masking does not introduce spurious correlation between two maps. In this approach, we split the particle image signal in real space into two components: 1) the real part, which is the signal used in traditional calculations; and 2) the imaginary part, which is the signal present only when Ewald's sphere calculation is considered. If the two components are used in reconstruction separately, they are independent of each other. Then, the FSC is calculated between two maps derived from the real part of all particles for one map and from the imaginary part of all particles for the second map. Typically, reconstruction from the real component is used to refine particle orientations, introducing bias into subsequent reconstruction based on the real component. However, reconstruction based on the imaginary component is not biased by the refinement using only the real component. This way, reconstruction from the imaginary component can be used as an unbiased reference for validating reconstruction from the real component. This validation works even at relatively low resolution (10 Å data if the number of particles is large), despite relying on Ewald's sphere curvature. The byproduct of this validation is the determination of the structure's handedness, which only requires medium resolution data. In cryoEM 3D reconstruction, noise is spread over the entire 3D space, while the signal is present only within the molecular mask. The signal-to-noise ratio (SNR) within the molecular mask is the SNR of interest in data analysis. Any method that does not apply the molecular mask to noise underestimates the SNR. For instance, Fourier Shell Correlation (FSC) curves recalculated between two half-maps after the data are deposited in the PDB do not apply the molecular mask to noise, consequently providing a misleading impression of resolution. These approaches, relying on calculations without mask constraints on noise, try to avoid introducing biases within the mask area that could potentially arise at low resolution. Here, we present a method of FSC calculations in which masking does not introduce spurious correlation between two maps. In this approach, we split the particle image signal in real space into two components: 1) the real part, which is the signal used in traditional calculations; and 2) the imaginary part, which is the signal present only when Ewald's sphere calculation is considered. If the two components are used in reconstruction separately, they are independent of each other. Then, the FSC is calculated between two maps derived from the real part of all particles for one map and from the imaginary part of all particles for the second map. Typically, reconstruction from the real component is used to refine particle orientations, introducing bias into subsequent reconstruction based on the real component. However, reconstruction based on the imaginary component is not biased by the refinement using only the real component. This way, reconstruction from the imaginary component can be used as an unbiased reference for validating reconstruction from the real component. This validation works even at relatively low resolution (10 Å data if the number of particles is large), despite relying on Ewald's sphere curvature. The byproduct of this validation is the determination of the structure's handedness, which only requires medium resolution data.

Improving dense conditional random field for retinal vessel segmentation by discriminative feature learning and thin-vessel enhancement

The analytic expression for the nonlinear magnetic susceptibility of a thin disk derived from Bean’s model with a uniform J C for predicts a peak-peak ratio of 7.1 (5.4) for the real (imaginary) components of the third harmonic χ 3 susceptibility. Our measurements show a peak-peak ratio closer to 1.1 in the real component and a noise limited lower estimate of ≈ 7 in the imaginary component. The anomalous third harmonic can be explained by the presence of a geometrical surface barrier for flux entry and exit which we have included as a small region of enhanced current density, J C,G, close to the surface of the disk. The geometrical surface current density is predicted to be approximately equal to J C,G ≈ 2B C1/μ 0 d. In 2G HTS tapes with d ≈ 1μm and B C1 ≈ 5mT this is J C,G ≈ 8GAm−2. We have measured both J C,B and J C,G using the nonlinear susceptibility of a Superpower tape without artificial pinning in fields orthogonal to the tape up to 35T. We find a geometrical surface current of J C,G ≈ 10GAm−2 and an average critical current density J C which is in good agreement with transport measurements performed on material from the same reel of tape.

Face detection is important for face localization in face or facial expression recognition, etc. The basic idea is to determine whether there is a face in an image or not, and also its location, size. It can be seen as a binary classification problem, which can be well solved by support vector machine (SVM). Though SVM has strong model generalization ability, it has some limitations, which will be deeply analyzed in the paper. To access them, we study the principle and characteristics of the Multiple Kernel Learning (MKL) and propose a MKL-based face detection algorithm. In the paper, we describe the proposed algorithm in the interdisciplinary research perspective of machine learning and image processing. After analyzing the limitation of describing a face with a single feature, we apply several ones. To fuse them well, we try different kernel functions on different feature. By MKL method, the weight of each single function is determined. Thus, we obtain the face detection model, which is the kernel of the proposed method. Experiments on the public data set and real life face images are performed. We compare the performance of the proposed algorithm with the single kernel-single feature based algorithm and multiple kernels-single feature based algorithm. The effectiveness of the proposed algorithm is illustrated. Keywords: face detection, feature fusion, SVM, MKL

The selection of motion feature directly affects the recognition effect of human action recognition method. Single feature is often affected by human appearance, environment, camera settings and other factors, and its recognition effect is limited. This paper propose a novel action recognition method by using selective ensemble learning, which is a special paradigm of ensemble learning. Moreover, this paper presents a fast and efficient action description feature and a novel recognition algorithm. Robust discriminant mixed features are learnt from behavioral video frames as behavioral descriptors, The recogniton algorithm using selective ensemble learning can achieve fast classification. Experimental results show that the proposed method achieves ideal recognition results on the self-built indoor behavior data set and public data set.