Visual evoked potential (VEP)-based brain–computer interfaces (BCIs) are widely employed but suffer from a fundamental usability barrier: reliance on perceptible low-frequency flicker that is uncomfortable for the user. Addressing this limitation at its root requires substantially attenuating perceived flicker, an approach feasible via high-frequency stimulation above the critical flicker fusion threshold, yet commonly assumed to be limited in signal-to-noise ratio and incompatible with the code-modulated visual evoked potential (c-VEP) paradigm. To bridge this gap, we introduce the code-modulated visual evoked potential paradigm with an imperceptible carrier (I-c-VEP), in which an imperceptible high-frequency carrier (60 Hz) is intermittently activated according to a pseudorandom code sequence. We demonstrate that the I-c-VEP paradigm, despite its low perceptual salience, is reliably decodable in real-time, evidenced by a mean online accuracy of 92.5% across 20 participants. This performance is achieved with a substantially lower flashing sensation than that of conventional c-VEP BCIs, translating to a mean comfort rating 3.85 points higher (on a six-point scale) compared to conventional c-VEP. By addressing the fundamental usability challenge of VEP-based BCIs, this work could open the door to genuinely user-friendly systems with broader practical applicability.
Periapical radiographs are widely used in clinical dentistry to assess periodontal bone loss percentage, typically relying on the judgment of dental professionals. However, current assessment methods face two significant challenges: 1) subjective interpretation is prone to misdiagnosis due to image quality issues and superimposition and is time-consuming; 2) most existing computer-aided systems are semiautomated, capable only of classifying the severity of periodontal bone loss, while percentage quantification still requires manual effort. We propose a fully automated periodontal bone loss percentage assessment framework based on deep learning and image processing techniques to address these limitations. The framework incorporates a single/multiroot tooth segmentation algorithm with AI-based assistance to isolate individual root regions for apical-level analysis. YOLOv11s integrated with preprocessing techniques achieves 98.8% accuracy in single-tooth detection, ensuring precise dental and bony structures identification. A segmentation model and a postprocessing algorithm then quantitatively evaluate periodontal bone loss based on alveolar crest and cementoenamel junction landmarks. This approach reduces diagnostic errors caused by superimposition and maintains over 90% diagnostic accuracy, outperforming current semi/fully automated methods by at least 5%. Clinically, the framework attains an average accuracy of 97.61% with a minimal error of 0.011. The proposed framework reduces diagnostic errors, and provides a reliable, objective tool to enhance efficiency and quality in patient care.
Audio–visual speech enhancement (AVSE) aims to improve speech quality and intelligibility in noisy environments by leveraging complementary visual cues. However, existing AVSE approaches typically rely on shallow or single-stage fusion strategies, which are insufficient for modeling hierarchical cross-modal dependencies and often limit robustness in real-world human–machine interaction (HMI) scenarios. To address this limitation, we propose a hierarchical audio–visual fusion network (HAFNet), an end-to-end time-domain AVSE framework built upon a convolutional timedomain audio separation network (Conv-TasNet) backbone. HAFNet incorporates a multilevel audio–visual fusion module composed of stacked cross-attention layers, enabling progressive and bidirectional interaction between audio and visual modalities. This hierarchical design allows the neural model to dynamically refine multimodal representations across multiple levels of abstraction, facilitating more effective exploitation of complementary cues under severe noise conditions. The proposed HAFNet is evaluated on the GRID-CHiME, LRS3, and COG-MHEAR challenge datasets using standard objective metrics. Experimental results demonstrate that HAFNet achieves significant improvements in highly noisy environments (SNR $\leq -12$ dB), attaining a short-time objective intelligibility of 72.58%, a perceptual evaluation of speech quality of 2.56, and an scale-invariant signal-to-distortion ratio of 7.80 dB, outperforming state-of-the-art benchmarks. These findings highlight the effectiveness of hierarchical multimodal fusion for robust speech enhancement and its potential to enable reliable communication in next-generation HMI systems.
The biggest challenge in the current research is the smooth planning of robots in dynamics environment. To cater the challenge, a map-based DAYINDI-AI Technique is developed and implemented to control humanoid and mobile robots in complex dynamic environment. Intelligent robot navigation is carried out using novel graphical map-based DAYINDI AI technique. The humanoid and mobile robots use ultrasonic, infrared, and image sensors to map workspaces and detect obstacles. These sensor inputs are represented as graphical arc segments in the proposed technique. The percentage of pilot radius (PR) has been calculated based on the sensory information. The robot’s threshold value has been calculated according to the percentage of the PR. The grand output obtained from the developed approach decides the steering angle during navigation in search space. The heading angle depends upon the goal and position of dynamic and static obstacles. Several environmental conditions have been taken for performing real-time experiments and simulation exercises using the proposed technique. For authentication of proposed technique, DAYINDI AI technique is compared with existing researches and significant improvements are observed. This approach is applicable for autonomous navigation and control of various robotic agents.
For individuals with limb impairments, prosthetic hands controlled by surface electromyography (sEMG) help restore hand movement and grasping ability. These systems are also an integral part of human–machine interfaces, enabling assistive robots to perform manipulation tasks. However, sEMG sensors face certain limitations in comprehensively capturing muscle activity due to their sensitivity to external interferences, and intra/intersubject variability. Classical dimensionality reduction techniques often underperform in such conditions, particularly when the intrinsic spatial structure of the data is overlooked. This article introduces a novel subspace projection method, named robust possibilistic neighborhood graph-aided orthogonal locality preserving projection (RPNG-OLPP), aimed at enhancing pattern discrimination under uncertain and nonlinear manifolds. Unlike traditional methods that depend solely on geometric proximity or class labels, RPNG-OLPP constructs a data-adaptive graph using possibilistic similarity estimation to reflect both local affinity and uncertainty-aware neighborhood structure. To enhance robustness, an $\ell _{2,1}$ norm-based sparsity constraint is introduced, suppressing features with poor discriminative power in the final representation. Minimizing this norm forces the rows of the projection matrix to shrink, effectively eliminating feature outliers and thereby enhancing the reliability of the similarity graph. To ensure feature decorrelation and numerical stability, the framework employs orthogonal regularization during the projection learning. Extensive evaluations are conducted on both laboratory-acquired electromyography datasets and the benchmark NinaPro database under regular and noisy conditions. The results suggest RPNG-OLPP’s potential as a robust, interpretable subspace modeling tool for real-world applications in pattern recognition and sensor-driven intelligent systems, particularly within the broader context of human–machine systems.
