工程技术最新文献

Camouflaged Object Detection (COD), the task of identifying objects concealed within their environments, has seen rapid growth due to its wide range of practical applications. We propose a human-machine collaboration framework for COD, leveraging the complementary strengths of computer vision (CV) models and noninvasive brain-computer interfaces (BCIs). Our approach introduces a multiview backbone to estimate uncertainty in CV predictions, utilizes this uncertainty during training to improve efficiency, and defers low-confidence cases to human evaluation via RSVP-based BCIs during testing for more reliable decision-making. Evaluated on the CAMO dataset, our framework achieves state-of-the-art results with an average improvement of 4.56% in balanced accuracy (BA) and 3.66% in the F1 score. For the best-performing participants, improvements reached 7.6% in BA and 6.66% in the F1 score. Training analysis showed a strong correlation between confidence and precision, while ablation studies confirmed the effectiveness of our training policy and human-machine collaboration strategy. This work reduces human cognitive load, improves system reliability, and provides a foundation for advancements in real-world COD applications and human-computer interaction. Our code and data are available at: https://github.com/ziyuey/Uncertainty-aware-human-machine-collaboration-in-camouflaged-object-identification.

The clinical manifestations of early-stage parkinsonian syndromes overlap, making accurate differential diagnosis crucial yet challenging. This study aimed to develop a system for automated differentiation of idiopathic Parkinson's disease (IPD) from progressive supranuclear palsy (PSP) and corticobasal syndrome (CBS). Our sample included clinical data and T1-weighted magnetic resonance imaging from 50 IPD, 47 PSP, and 38 CBS patients. We introduced an atlas-based approach to extract shape features from subcortical regions in each subject's native coordinate image space. The surface thickness and folding parameters were also extracted from cortical regions. A statistical analysis was conducted to identify regions with significant differences in the extracted features, followed by the employment of a feed-forward neural network to distinguish these patients. Significant structural differences were observed in several regions, including the thalamic nuclei, basal ganglia, midbrain, cerebellum, cingulate cortex, and insula. Using only cortical surface features, our diagnostic model outperformed the model that relied solely on subcortical shape features. However, the classifier achieved its best predictive performance when incorporating features from both cortical and subcortical structures, yielding an accuracy of 86.1% in a multi-class classification system and 96.1% for distinguishing IPD from PSP and CBS, as well as an accuracy of 94.2% for classifying CBS versus PSP in a two-class classification system. Our findings underscore the significance of cortical morphological patterns and demonstrate that the proposed methodology could potentially serve as an automated diagnostic system in clinical settings.

Supplementary information: The online version contains supplementary material available at 10.1007/s11571-025-10402-2.

Place cells and head direction cells in the rodent brain encode spatial position and orientation, forming the neural basis for navigation and cognitive map construction. Inspired by these mechanisms, RatSLAM simulates their roles to achieve biologically inspired visual SLAM. However, traditional RatSLAM struggles with robust feature extraction in visually complex or dynamic environments, where features may be unstable or non-distinct. To address this, we integrate the AKAZE algorithm into the RatSLAM framework. AKAZE combines accelerated techniques with nonlinear diffusion filtering to construct a multi-scale nonlinear scale space, enabling efficient extraction of robust, scale-invariant features across spatial scales. These features are incorporated into RatSLAM's local view module to improve loop closure detection and mitigate odometry drift. Traditional evaluation approaches rely on real-time pose trajectories and cannot evaluate the trajectories based on the fully optimized experience maps, leading to inaccurate mapping performance assessments. Thus, we further propose a novel Ray-Based Map Metric Error Evaluation Method, which can directly compare the final experience maps generated by RatSLAM. Experiments on the KITTI dataset demonstrate that, compared with both ORB-RatSLAM and the ORB-SLAM3, the proposed AKAZE-RatSLAM achieves higher loop closure recall and mapping accuracy while maintaining a lightweight computational profile. In particular, CPU and memory measurements show that AKAZE-RatSLAM requires significantly less computational resources than ORB-SLAM3, confirming its suitability for real-time deployment on resource-limited robotic platforms. Furthermore, neuro-inspired analyses reveal that the pose cell network exhibits spatially localized and direction-selective firing patterns analogous to hippocampal place cells and head direction cells in rodents. Specifically, cells along the same row encode adjacent spatial regions, forming continuous place-field-like activations, whereas cells in the same column show distinct preferred orientations, indicating directional tuning. These biological characteristics confirm that the proposed AKAZE-RatSLAM not only enhances mapping performance and efficiency but also preserves the neurobiological plausibility of spatial representation, advancing the development of brain-inspired visual SLAM systems.

In surgical procedures, surgeons can suffer from spontaneous hand tremors that can affect the accuracy of surgical robots. Therefore, it is necessary to measure and model the tremor signal by sensors to suppress hand tremor. This paper proposes a prediction method based on deep learning that integrates long-term and short-term features to achieve this goal. The long-term features of tremor signals are extracted using a bidirectional Long-short-term memory network, while the short-term features are extracted using a Temporal Convolutional Network. By integrating the long-term and short-term characteristics of tremor signals, this approach provides rich temporal information for signal estimation. In addition, genetic algorithm is used to obtain the optimal time step-size to fully explore the temporal correlation of signals, and an end data compensation strategy is adopted to ensure that the tremor filtering covers the entire process. The performance of the proposed method is evaluated by training and testing on the same dataset as other methods, and conducting suture experiments in a virtual surgical environment. The results show that our proposed model is superior to the existing methods, effectively reducing the tremor signals estimation error. This method can provide better tremor estimation and compensation performance, effectively suppressing the hand tremors and improving the surgical accuracy.

EEG signals are widely used in emotion recognition due to their capability for objective emotional state quantification. However, despite containing abundant frequency and spatial information, researchers continue to face challenges in extracting fine-grained discriminative features from these signals. We develop SC-SDT (Spectral Convolution-Spatial Differential Transformer), a novel framework that jointly models spectral and spatial characteristics through an integrated convolutional and transformer architecture. First the model is equipped with a Spectral Feature Embedding module that employs a sequential group-pointwise convolutional network. This enables the dynamic capture of both local spectral patterns within bands and global interactions across the frequency spectrum. Subsequently, a Spatial Feature Extraction module is designed to simultaneously mitigate attention noise and optimize functional connectivity mapping across EEG channels through its core differential attention mechanism. Finally, to enhance model robustness against inter-subject variability, we introduce supervised contrastive loss that explicitly enforces subject-invariant feature representations while preserving class discriminability. Employing a subject-independent experimental paradigm, we rigorously evaluated the proposed SC-SDT model on SEED, SEED-IV, and DEAP datasets to assess cross-subject generalization capabilities. Experimental results demonstrate that SC-SDT achieves competitive emotion classification performance by effectively modeling spectral-spatial neural signatures. Our analysis of its key components further reveals that the model not only pioneers the application of differential attention in EEG, but also offers a methodological foundation for efficient spectral-spatial feature extraction. The code for this paper is accessible at https://github.com/apolloCoder-byte/SC-SDT.

Numerous studies in the bilingual literature have shown that cognitive control adapts to several factors related to second language (L2) learning. However, whether third language (L3) learning influences cognitive control remains underexplored. In this longitudinal study, we analyzed behavioral performance and functional magnetic resonance imaging (fMRI) data among Chinese-English bilinguals at resting-state and during a flanker task both prior to English (L2) or Japanese (L3) learning and one year later. During brain resting-states for these same learners, we conducted a correlation analysis between language exam scores and functional connectivity strength of resting-state data after one year of study. The connectivity between the left anterior cingulate cortex (ACC) and the left precuneus was positively correlated with English listening performance, while the connectivity between the right supramarginal gyrus (SMG) and the right inferior parietal lobe (IPL) was negatively correlated with English oral performance. The behavioral results from the flanker task showed that after one year of L2 learning in a classroom setting, a significantly smaller flanker effect emerged among Chinese-English bilinguals. Moreover, brain imaging revealed that incongruent flanker trials elicited greater activation of the left superior frontal gyrus (SFG) than congruent trials. These behavioral and neural patterns were not found among Chinese-English bilinguals who had studied Japanese for one year. Taken together, these findings suggest that cognitive control adapts to L2 learning, but appears to be unaffected by L3 learning.

Supplementary information: The online version contains supplementary material available at 10.1007/s11571-025-10397-w.

Abnormal τ and β-amyloid (Aβ) deposition in the brains of patients with Alzheimer's disease (AD) is significantly associated with cognitive decline. This abnormal deposition has been reported to be linked to increased excitatory and inhibitory time constants in neural circuits. In this paper, we focus on three typical electroencephalography (EEG) slowdowns clinically reported in association with AD, including decreased dominant frequency, decreased α rhythmic activity, and increased δ + θ rhythmic activity. Firstly, we demonstrate that changes in excitatory time constant, inhibitory time constants, and synaptic connection strength can induce the phenomenon of EEG slowdowns in early AD. Then, we are interested in the regulation of AD by traditional deep brain stimulation (DBS) and emerging optogenetic stimulation. High-frequency, high-pulse width, and high-amplitude DBS are more effective in reversing brain rhythm in AD, supporting the experiment that cortical high-frequency DBS may be an effective therapeutic way for dementia-related diseases. In particular, as a modification of traditional DBS, we find that oscillatory bursty stimulation can compensate for the shortcomings of DBS at low amplitude. However, it is physiologically difficult to target inhibitory interneurons with conventional electrical stimulation. Optogenetics is able to precisely stimulate pyramidal neurons and inhibitory interneurons observed in animal experiments. Our numerical results indicate that medium and low-frequency stimulation can better eliminate AD pathology. It should be noted that stimulation of inhibitory interneurons requires greater light intensity than stimulation of pyramidal neurons. Finally, we propose two optimization intermittent optogenetic stimulation protocols. These modeling results can reproduce some experimental phenomena and are expected to reveal the underlying pathological mechanisms and control strategies associated with cognitive dysfunction such as AD.

Brain-controlled wheelchair (BCW) technology enables direct wheelchair control via brain-computer interfaces (BCIs), eliminating the need for physical limb interaction. Motor imagery-based BCIs (MI-BCIs) are widely used in non-invasive BCIs due to their ability to provide intuitive neural control without external stimuli. However, developing a BCW system based on MI-BCIs remains challenging, particularly in achieving reliable multi-class classification accuracy.To address this challenge, this study proposes an advanced feature extraction algorithm to enhance MI-BCI performance using a custom-built five-class MI-EEG dataset. The proposed method, EHT-CSP, integrates Ensemble Empirical Mode Decomposition Hilbert-Huang Transform (EEMD-HHT) with Time-Frequency Common Spatial Pattern (TFCSP). Specifically, it extracts marginal spectrum entropy and energy spectrum entropy via EEMD-HHT. It then combines these features with TFCSP-derived feature vectors to improve feature discrimination. The Light Gradient Boosting Machine is then employed for classification. The proposed MI-BCI system is evaluated through both offline analysis and real-world BCW obstacle avoidance experiments. Results demonstrate that the algorithm achieves an average classification accuracy of 78.45%, with all participants successfully completing BCW navigation tasks. In this study, LightGBM and EHT-CSP are compared with other algorithms respectively, and it is verified that the proposed model is superior to the existing models.

Pediatric exoskeletons have the potential to aid the walking of children with neuromuscular conditions such as crouch gait. However, current exoskeleton devices often rely on bulky batteries and motors. Recent developments in 3D-printing technologies now allow the construction of lightweight yet stiff parts that are easy to customize and use for pediatric applications. We present the mechanical design of a 3D-printed and spring-powered knee exoskeleton for gait assistance. The device had a mass of ∼1.25 kg per leg and provided a knee extensor moment during the stance phase of gait, simulating the spring-like behavior of the knee. Conversely, the exoskeleton provided no resistance during swing to allow free motion of the joint. To validate the device, we recruited two neurologically intact children to walk on a treadmill with and without the exoskeleton while we recorded kinematics, kinetics, and muscle activity data. Our exoskeleton generated knee extensor moments proportional to its angular excursion and had a peak mean moment of ∼0.1 N·m/kg during stance. Kinetic data showed that subjects decreased their biological knee moment and joint spring-like behavior to compensate for the added exoskeleton moment and stiffness, respectively. We ultimately show that the device is robust and capable of generating extensor moments comparable to devices used to assist the knee in children with crouch gait.