Emotions are pivotal in influencing students’ learning experience, academic achievement, and general well-being. Fear and anxiety, especially, are hindrances to cognitive processes like memory, concentration, and problem-solving, usually evoked by high-stakes academic contexts. Conventional self-reported questionnaires and psychological tests employed for fear measurement are plagued by subjectivity, response bias, and a lack of ability to reflect real-time emotional fluctuations. To overcome these constraints, this research presents PsyVisionNet, a multi-modal deep learning platform that incorporates computer vision-based facial emotion analysis and psychological parameters (heart rate, stress level, and anxiety scores) for real-time fear recognition among students. The proposed model utilizes a pre-trained ResNet-18 to extract facial features associated with fear and a Multi-Layer Perceptron (MLP) to process physiological information. These characteristics are integrated in a decision-making layer for categorizing students into “Phobic” or “Non Phobic” classes. PsyVisionNet yields 92.5
In this paper, a new adversarial imitation learning algorithm is proposed based on Weighted Wasserstein distance (WWDAIL). This algorithm employs an enhanced proximal optimization strategy to improve the efficiency and stability of parameter updates. Furthermore, the weighted Wasserstein distance is utilized as a metric for assessing policy differences, thereby increasing the flexibility and dynamism of the measurement. Furthermore, by refining the reward function derived from the discriminator's output, the reward collection capability and efficiency are significantly enhanced. Experimental results in the Mujoco simulation environment have demonstrated that the proposed algorithm exhibits substantial stability in continuous control tasks and markedly improves the agent's reward collection ability and efficiency.
A quantum multi-classification network based on parameterized quantum circuits is proposed for handwritten digit image classification tasks. Classical 28 × 28 pixel images are downsampled to 4 × 4 resolution, and the resulting 16 classical data points encoded into probability amplitudes of 4-qubit quantum states. These quantum states are input into a quantum state preparation circuit. The parameterized quantum circuit, composed of rotation and entanglement layers with trainable parameters, is designed to process the input quantum states. Projective measurements are performed on the output quantum states to obtain measurement vectors, which are transformed into output value vectors corresponding to One-hot labels through the Softmax function. The quantum 4-class classification network is taken as an example to introduce in detail. We also extended to 5–10 classification tasks.
Stock price fluctuations are characterized by randomness and complexity, making accurate prediction a challenging task in financial markets. This paper proposes a dual-modal stock price change rate forecasting framework based on pretrained models (DMFPre) to improve the accuracy of stock price change rate forecasting. The framework consists of two main modules: on one hand, it utilizes a pretrained language model combined with company-related textual data to predict the weekly stock price change rate range. The prediction accuracy is closely related to the design of the prompts, and manually designed prompts not only have significant room for optimization but also involve time-consuming and labor-intensive adjustments. To address this, the paper introduces an automatic prompt optimization method that automatically updates the prompts through algorithms. On the other hand, the framework employs a pretrained time series forecasting model (PTSFM) to predict the daily price change rate trend based on historical stock trading data. Finally, by integrating the prediction results from both modules, the framework generates more accurate stock price change rate forecasting. Experimental results show that on the CSI 100 constituent stock dataset, the DMFPre framework significantly outperforms traditional single-modal time series forecasting models in terms of prediction accuracy.
In this paper, we propose adaptive control schemes for the universal function projective synchronization of delayed Chen system and delayed Lorenz system with unknown parameters. We first consider the situation where a vector is used as the reference system. Adaptation control scheme is used for the unknown parameters. Then, we consider the situation where an uncertain Liu system is used as the reference system. We add switching control terms in the control laws for external disturbances. To alleviate the chattering, fuzzy systems are used in the controllers. Simulation results demonstrate the effectiveness of the proposed methods.
Hyperspectral imaging provides a powerful means of material identification by capturing detailed spectral information across a broad range of wavelengths. However, the accurate classification of materials remains challenging due to spectral mixing and the lack of ground truth data. This paper proposes a robust approach for hyperspectral endmember extraction and material identification by integrating the N-FINDR algorithm with spectral information divergence (SID)based spectral library matching. The proposed method enhances segmentation accuracy by identifying pure spectral signatures (endmembers) and associating them with actual materials using the ECOSTRESS spectral library. Experimental validation using the Pavia University hyperspectral dataset demonstrates the effectiveness of this approach in extracting, identifying, and segmenting endmember materials with high precision. The results underscore the advantages of combining geometric-based endmember extraction with spectral matching techniques to improve hyperspectral image analysis.
The output tracking problem is an important topic in the field of system control, and the pole placement based Zhang neurodynamics (PP-based ZN) method is an effective way to solve this problem. In this paper, for the convenience of digital hardware implementation, two time-discretization methods, namely, Euler discretization and time-invariant linear system (TILS) special discretization, are applied to conduct discrete control and its simulations based on the PP-based ZN output tracking controller. In numerical experiments, both discrete forms show high precision for different sampling interval gaps and desired tracking trajectories (i.e., paths). Besides, through the discussion of simulation results, the similarities and differences between the two discrete forms are compared.
Transformer-based models have demonstrated state-of-theart results in the field of image super-resolution. However, we observe that such methods sometimes suffer from overly smooth structural reconstruction and blurred details, indicating that the potential of Transformers has not yet been fully exploited in existing networks. To leverage more prior information, this paper proposes a novel Global Cross Attention Transformer (GCAT) algorithm. This algorithm introduces external prior information by incorporating a cross-attention mechanism alongside the original self-attention mechanism. Furthermore, to better establish the model, we apply cross-attention across all Transformer modules to enhance the model capability for complex mapping. Extensive experiments demonstrate the efficacy of the proposed architecture, with the overall approach exceeding the performance of current state-of-the-art methods.
Blind source separation (BSS), also known as audio signal separation, is the process of extracting individual sound sources from an audio mixture. This technique has widespread applications in audio signal processing, speech enhancement, and other related fields. The present study investigates the separation of sound signals contaminated by additive white Gaussian noise (AWGN). Due to the presence of such noise, accurately retrieving individual signals becomes a significant challenge. This work employs independent component analysis (ICA) to address the separation problem. A mixing matrix, incorporating down-sampled signals, is dynamically generated using AWGN prior to signal mixing. To further suppress noise in the combined signal, a Fast Fourier Transform (FFT)-based de-noising method is applied. Subsequently, the Inverse Fast Fourier Transform (IFFT) is used to reconstruct the separated signals. Compared with conventional approaches, the proposed framework enhances noise removal and improves the auditory quality of the recovered signals, rendering them perceptually close to the originals.
Ship course tracking (i.e., ship heading tracking) system is one of the key systems for maritime navigation. Due to the non-linearity of the ship motion and environmental disturbances, it is relatively challenging to realize the effective control of the ship heading. Moreover, compared with continuous-time systems, discrete-time systems are preferable in practical applications. In this paper, on the basis of continuous ZN (Zhang neurodynamics) PD (proportional and derivative) and PID (proportional, integral and derivative) controllers, discrete ZN-PD and ZN-PID controllers for the ship course tracking are further designed and simulated. Specifically, the Eulerian discretization method is used, which uses numerical differentiation to approximate the derivatives to discretize the continuous controllers into discrete controllers. Finally, the simulation results confirm that both controllers provide effective heading tracking, with the ZN-PD controller exhibiting faster response and the ZN-PID controller effectively improving response speed and providing better tracking for complicated dynamic inputs.
This paper investigates the distributed fault-tolerant consensus problem for nonlinear multi-agent systems (MASs). A novel distributed fault-tolerant control protocol is proposed under a zero-sum differential game framework, where the consensus problem is reformulated as a minimax optimization between the control inputs of agents and the actuator faults through a local cost function. A critic neural network is trained online to solve the coupled Hamilton-Jacobi-Isaacs (HJI) equation, where the optimal control and upper bound of fault compensation are simultaneously derived from the Nash equilibrium condition. Leveraging the Lyapunov stability theorem, it is rigorously proved that the designed distributed fault-tolerant consensus control law guarantees the uniform ultimate boundedness (UUB) of the closed-loop systems. Simulation results validate the effectiveness of the present method.
This paper focuses on the trajectory tracking control of an underactuated AUV in the vertical plane. First, an error analysis is conducted for the underactuated AUV in the vertical plane, and a Disturbance Observer (DO) is designed to estimate external disturbances, while a Radial Basis Function Neural Network (RBFNN) is employed to approximate nonlinear terms. Subsequently, controllers are designed based on Nonsingular Terminal Sliding Mode Control (NTSMC) and Nonsingular Fast Terminal Sliding Mode Control (NFTSMC), respectively. Finally, the stability of the controllers is verified using the Lyapunov function. Simulation results demonstrate that controllers effectively improve trajectory tracking accuracy and enhances the robustness of the system.
Swine behavior monitoring is pivotal for precision livestock farming, and posture recognition remains challenging. We propose a hybrid LSTM-CNN algorithm utilizing neck-mounted inertial sensors to classify four typical postures standing, feeding, lying, and active movement. By extracting time-domain features such as signal magnitude and variance from triaxial acceleration data, the architecture integrates CNN’s spatial pattern learning with LSTM’s temporal sequence modeling. The model achieves 73.2
With the progress of data acquisition technologies, semisupervised multi-view learning (SSML) has become a research hotspot in machine learning. Although graph convolutional network (GCN) have garnered significant attention in the domain of SSML due to their unique ability to propagate label signals through graph structures, existing methods still face limitations in feature fusion, which affects the performance of multi-view classification. To address this challenge, we propose a novel framework called Graph-based Consistent Feature Fusion (GCFF). The framework first employs view-specific GCNs to extract features from each view. Subsequently, it fuses these features through an adaptive weighting mechanism and optimizes the fused features using a consistency feature constraint loss, thereby significantly enhancing the quality of the fused features. Finally, a view-common GCN integrates the fused graph structure and consistent features to propagate label signals efficiently, yielding better classification performance. Experiments on four datasets demonstrate the superior performance of GCFF.
Robotic manipulators invariably encounter noise disturbances in practical applications. In the research on motion planning of robotic manipulators, the impact of noise and the optimization of control parameters are often overlooked, which can result in suboptimal noise resistance and prevent robots from successfully completing their tasks. To address this issue, this paper proposes a fuzzy logic-based noisesuppression scheme for motion planning of robotic manipulators. This scheme builds upon the pseudo-inverse method by incorporating error feedback and integral gain to suppress noise. Furthermore, the designed fuzzy rules optimize the control parameters, ensuring effective noise suppression during the robot's motion planning phase. Finally, through simulations and comparisons, the feasibility and superiority of the proposed fuzzy-rule-based noise-resistant motion planning scheme are validated.
In response to the rapidly evolving demands of modern manufacturing, traditional manual inspection of metal surface defects in industry can no longer meet the requirements for fast and high-precision detection. To address this challenge, this paper presents an enhanced efficient detection algorithm based on YOLOv8. The model improves multi-scale feature perception and computational efficiency by incorporating Dual Convolution (DualConv) and Large Kernel Separable Attention (LSKA) into the neck and backbone, respectively. Additionally, wavelet pooling is introduced to optimize the sampling process. These enhancements reduce the model’s parameters while further increasing its accuracy. Compared to the original YOLOv8, GFLOPs are reduced by 14.8
In this paper, we propose a multiple-order time-delay Zhang neural dynamics (MOTDZND) model for handling the tracking control problem of the chaotic system with mixed input. The key technique of the MOTDZNN model is to approximate multiple-order derivatives of the desired path by using the backward finite difference (BFD) rules. By adopting a group of BFD rules, we develop a MOTDZND model, whose truncation error is square-form, for handling the tracking control problem of the chaotic system with mixed input. A group of simulations verifies that the proposed model is validity and consist with the theoretical square-form truncation error.
Fundus images and optical coherence tomography (OCT) provide complementary diagnostic information from retinal surface structures and deep tissue tomograms, respectively, playing a critical role in early screening of blinding diseases such as diabetic retinopathy and glaucoma. However, the differences in imaging principles between the two modalities (e.g., color texture of fundus images vs. tomographic grayscale features of OCT) lead to significant heterogeneity in feature spaces. Traditional single-modal models (e.g., ResNet-based single-image classifiers) often suffer from incomplete feature representation in complex lesion recognition due to the lack of multimodal information interaction. To address this, this paper proposes an early-stage feature fusion framework based on dual MobileNetV3 networks, which extracts discriminative features through modality-specific networks and enhances pathological region focusing using a self-attention mechanism. The framework achieved a 98.9
Face recognition (FR) remains a critical component in security and authentication applications, requiring models that balance accuracy and computational efficiency. Traditional FR models rely on MLP-based classification heads, which serve as the standard approach for feature representation, but it remains an open research question as to which variant offers the optimal balance between model capacity and feature discriminability. In this paper, we introduce KANFace, a novel framework that integrates Kolmogorov-Arnold Networks (KANs) into face recognition architectures by replacing the conventional MLP-based head with a KANs module driven by adaptive, learnable B-spline activations. Evaluated across diverse datasets, including LFW, CFP-FP, and IJB-C, KANFace achieves state-of-the-art (SOTA) performance on CFP-FP, surpassing the baseline by 3
In the process of social opinion evolution, the delay in information transmission and the asynchrony of individual updates can lead to lag errors, which affect the accuracy of opinion evolution analysis. To address this issue, this paper proposes a new competitive model based on.k-winners-takes-all (k-WTA) network, which is used to describe and analyze the dynamic evolution of opinions in social networks. This model not only effectively eliminates lag errors but also takes into account the potential weight-unbalanced communication topology within real-world social networks. Furthermore, it supports distributed opinion exchange and evolution, enhancing the model's applicability in real-world scenarios. Finally, simulations further validate the effectiveness and feasibility of this model.
