We present EMG-Adapt, a novel few-shot prototype adaptation framework designed to enhance the robustness and data efficiency of electromyography (EMG)-based gesture recognition. By integrating the representational power of prototype learning with the rapid adaptation capabilities of meta-learning, our framework introduces several technical novelties. These include a cepstrum coefficient average feature extraction method that reduces sensitivity to noise and variations, a deep prototype learning method based on hybrid loss functions for both discriminative classification and embedding space structure, and a meta-learning strategy for efficient prototype update with minimal labeled examples. Our integrated approach significantly improves few-shot gesture recognition performance, requiring substantially less calibration data than conventional methods. Extensive experiments on five public EMG datasets demonstrate state-of-the-art performance in cross-session and cross-user generalization scenarios, while maintaining computational efficiency. This work represents a significant advancement towards practical, user-friendly, and scalable EMG-based human-computer interfaces, with potential applications in prosthetics, assistive technologies, and virtual reality. Future research will explore self-supervised learning techniques and extend the framework to handle more gestures and online adaptation strategies for enhanced real-world robustness.

Large Vision-Language Models (LVLMs) with “multimodal distractibility,” where plausible but irrelevant visual or textual inputs cause significant drops in reasoning consistency and lead to unreliable outputs. This paper introduces a comprehensive framework to systematically diagnose, evaluate, and mitigate this critical challenge. We present three core components: the large-scale IR-VQA benchmark to surface these vulnerabilities across four paradigms; novel diagnostic metrics, Positive Consistency (PC) and Negative Consistency (NC), which move beyond standard accuracy to rigorously measure a model's reasoning stability; and the Relevance-Gated Multimodal Routing (RGMR) mechanism, a novel, lightweight module that proactively and dynamically filters distractions at inference time. Our experiments reveal that state-of-the-art models exhibit significant drops in consistency on IR-VQA. We demonstrate that finetuning on IR-VQA and deploying RGMR substantially improve model robustness where standard prompting fails. Our comprehensive analysis of model behaviors under different types of distractions and the underlying reasoning failures provides a clear path forward for developing more reliable multimodal systems.

Few-shot semantic segmentation aims to perform pixel-level recognition of novel classes from only a handful of annotated examples. Although cross-attention-based methods have shown strong performance, their quadratic computational complexity limits efficient modeling of high-resolution features. Recently, state space models (e.g., Mamba), with linear complexity, have provided a promising alternative. However, directly applying Mamba to few-shot segmentation still faces three key challenges: (1) convolutional backbones lack global coordinate awareness, resulting in insufficient spatial-context modeling; (2) cross-image Mamba fusion suffers from support-feature forgetting and intra-class discrepancy; and (3) fixed receptive fields cannot flexibly adapt to object-scale variations. To address these issues, we propose a Multi-Scale Hybrid Mamba Network. The core idea is to build a cooperative “enhance-fuse-refine” pipeline. First, a global grouped coordinate attention module explicitly models global dependencies along height and width before feature serialization, thereby improving spatial awareness. Second, a hybrid Mamba fusion mechanism is introduced to alleviate support-feature forgetting and intra-class discrepancy, enabling efficient cross-image interaction. Third, a multi-scale atrous fusion module employs parallel dilated convolutions to capture multi-scale context and refine boundary details. Experiments on PASCAL- 5i and COCO- 20i show that the proposed method outperforms existing state-of-the-art approaches while preserving linear computational complexity.

Evaluating time delay estimation performance is critical in reconfigurable intelligent surfaces (RIS)-assisted massive MIMO systems. However, classical performance bounds such as the CRB and BCRB are only locally tight. In this correspondence, we derive a simplified closed-form expression for the globally effective Ziv-Zakai bound (ZZB) by approximating the summation of discrete signal samples as an integral over continuous signals and applying the Fourier transform for frequency-domain analysis. Simulation results demonstrate that under uniform a priori distributions, snapshots don't affect the value of the ZZB. Furthermore, regardless of variations in the number of RIS passive elements and base station antennas, the ZZB consistently offers a tighter bound than CRB in the low SNR region.

This paper presents an error analysis of classical and learned Tikhonov regularization schemes for inverse problems. We first demonstrate, both theoretically and numerically, that using a fixed regularization parameter across varying noise levels-which is a common miss-specification in practice-has only a mild impact on the reconstruction error. As a special case, we then investigate scenarios where the true data resides in an unknown finite-dimensional subspace. Here, our results lead to an empirically supported strategy for estimating the unknown dimension based on numerical experiments. Finally, we examine the approach that motivated this study: a method where a sparsity-promoting term is learned from denoising tasks and subsequently applied to general inverse problems via a simple heuristic parameter selection. The corresponding error analysis is initially developed using classical concepts and subsequently refined through a more detailed investigation of the discretized setting.

Neural interface devices generate large amounts of data that must be transmitted to a host computer for real-time neural decoding. Low-power wireless protocols, such as Bluetooth low energy (BLE), have limited data throughput, creating challenges for high-fidelity data streaming. We propose a lossless compression method designed to optimize wireless transmission without compromising neural signal integrity. Our compression pipeline includes four key stages: data filtering, cross-channel transformation, predictive coding using an autoregressive (AR) model, and Huffman coding. The method adapts to dynamic neural signal properties by recalculating AR model coefficients at regular intervals, ensuring sustained compression efficiency. This approach achieves a compression ratio between 3.02 and 4.42 while maintaining low computational and memory complexity, making it suitable for implementation on low-power microprocessors. We validated this method on a prototype neural interface using the NRF52840 system-on-chip (SoC). The device streams 16 neural data channels at a 10-kHz sampling rate to a host via a BLE 5.2 data link. The system sustains a compression ratio of 3.5 and a data throughput of 220 kb/s, facilitating real-time, lossless transmission of neural signals with minimal latency. This innovation supports the development of wireless peripheral nerve interfaces for neuroprosthetic control, offering reliable real-time data transfer from neural interfaces to neural decoders over BLE.

Copyright protection of 3D point clouds is essential due to the increasingly widespread use of such media. Despite the effectiveness of recent watermarking techniques, ensuring resilience against various attacks while maintaining lossless quality and distinguishability remains a challenge. In this paper, we propose a novel zero-watermarking scheme with two complementary features to enhance performance under strong geometric and connectivity attacks while satisfying the other two criteria. Our approach involves extracting geometric features through spectral analysis of graphs built on spherical divided partitions to ensure distinguishability and resilience against common geometric attacks. Additionally, we incorporate a semantic feature extraction module using a dynamic graph convolutional neural network to withstand connectivity attacks, including cropping, simplification, and subdivision. Moreover, an attention-based fusion strategy is applied to bolster the robustness of the two features. Experimental results indicate that our scheme achieves a better trade-off between distinguishability and robustness with significantly lower false negative rates under the same false positive rate compared to state-of-the-art zero-watermarking methods against common geometric and connectivity attacks.

Upper extremity amputation, often necessitated by traumatic injuries, significantly impacts an individual's well-being. This paper addresses the critical challenges of deploying deep learning for real-time electromyography-based gesture recognition in prosthetic control: generalization across users and time, the personalization-generalization trade-off, and computational constraints. We propose Federated Electromyography (FedEMG), a novel Federated Prototype Learning (FPL) framework that leverages a prototype-based approach for efficient knowledge transfer and a unique adaptive personalization mechanism. Unlike existing Federated Learning (FL) methods, FedEMG balances global knowledge with user-specific adaptations, achieving high accuracy and personalization without sacrificing generalization. Furthermore, FedEMG utilizes a lightweight gesture detector in combination with an efficient neural network architecture optimized for resource-constrained devices, enabling real-time performance. Extensive evaluations on public and neural-prosthetic interface datasets demonstrate FedEMG's superior accuracy in intra- and inter-subject gesture recognition under various non-IID cases, while also highlighting its efficient resource utilization. FedEMG thus advances the field of upper-limb rehabilitation through improved and accessible prosthetic control.

Traditional object detection models often lose the detailed outline information of the object. To address this problem, we propose the Fourier Series Object Detection (FSD). It encodes the object's outline closed curve into two one-dimensional periodic Fourier series. The Fourier Series Model (FSM) is constructed to regress the Fourier series for each object in the image. Thus, during inference, the detailed outline information of each object can be retrieved. We introduce Rolling Optimization Matching for Fourier loss to ensure that the model's learning process is not affected by the sequence of the starting points of the labeled contour points, speeding up the training process. The FSM demonstrates improved feature extraction and descriptive capabilities for non-rectangular or elongated object regions. The model achieves AP50=73.3% on the DOTA 1.5 dataset, which surpasses the state-of-the-art (SOTA) method by 6.44% at 66.86%. On the UCAS dataset, the model achieves AP50=97.25%, also surpassing the performance indicators of the SOTA methods. Furthermore, we introduce the object's Fourier power spectrum to describe outline features and the Fourier vector to indicate its direction. This enhances the scene semantic representation of the object detection model and paves a new pathway for the evolution of object detection methodologies.

Deep saliency models, which predict what parts of an image capture our attention, are often like black boxes. This limits their use, especially in areas where understanding why a model makes a decision is crucial. Our research tackles this challenge by developing an explainable saliency (XSal) model that not only identifies what is important in an image, but also explains its choices in a way that makes sense to humans. We achieve this by using vision-language models to reason about images and by focusing the model’s attention on the most crucial information using a contextual prioritization mechanism. Unlike prior approaches that rely on fixation descriptions or soft-attention based semantic aggregation, our method directly models the reasoning steps involved in saliency prediction, generating selectively prioritized explanations clarify why specific regions are prioritized. Comprehensive evaluations demonstrate the effectiveness of our model in generating high-quality saliency maps and coherent, contextually relevant explanations. This research is a step towards more transparent and trustworthy AI systems that can help us understand and navigate the world around us.

Myoelectric prostheses, controlled by electromyographic (EMG) signals, suffer from significant inter-day performance degradation due to factors like muscle fatigue and electrode shift. Existing deep learning approaches often require extensive daily retraining, hindering real-world usability. This paper presents a novel transfer learning framework that mitigates this limitation by combining a pre-trained convolutional neural network (CNN) for robust feature extraction with a fine-tuned linear discriminator for efficient daily adaptation. Our methodology demonstrates superior inter-day generalizability and long-term stability compared to state-of-the-art methods on the Kanoga dataset while requiring substantially less daily calibration data.

The mixed Analog-to-Digital Converter (ADC) architecture is considered a promising solution in balancing the trade-off between high-resolution and low-resolution quantization over the hardware costs, power consumption, and transmission demands in massive multiple-input multiple-output (MIMO) systems. Meanwhile, the Direction of Arrival (DOA) estimation is a prerequisite for accurate beam processing in MIMO systems. Therefore, evaluating the DOA estimation performance in linear array architectures of mixed-ADC based MIMO systems is crucial. However, local bounds, such as the widely used Cramer-Rao Bound (CRB), only offer rigorous performance analysis of the estimator in the high signal-to-noise ratio (SNR) regime. In this paper, we derive a globally effective and closed-form Ziv-Zakai Bound (ZZB) to assess the DOA estimation performance of mixed-resolution quantization structures. We have also provided the CRB for DOA estimation with mixed-resolution quantization and included the classical MUSIC algorithm as a comparison. Additionally, we have analyzed the impact of different prior information, the numbers of snapshots and sensors, quantization bits and parameter settings on the ZZB. Simulation results show that the ZZB provides globally effective bounds under Gaussian and uniform distributions. In particular, in the low SNRs region, the ZZB offers a tighter and more effective bound than CRB and BCRB.

In recent years, significant strides in deep learning have propelled the advancement of electromyography (EMG)-based upper-limb gesture recognition systems, yielding notable successes across a spectrum of domains, including rehabilitation, orthopedics, robotics, and human-computer interaction. Despite these achievements, prevailing methodologies often overlook the intrinsic physical configurations and interconnectivity of multi-channel sensory inputs, resulting in a failure to adequately capture relational information embedded within the connections of deployed EMG sensor network topology. This oversight poses a significant challenge, impeding the extraction of crucial features from collaborative multi-channel EMG inputs and subsequently constraining model performance, generalizability, and interpretability. To address these limitations, we introduce novel graph structures meticulously crafted to encapsulate the spatial proximity of distributed EMG sensors and the temporal adjacency of EMG signals. Harnessing these tailored graph structures, we present Graph Convolution Network (GCN)-based classification models adept at effectively extracting and aggregating key features associated with various gestures. Our methodology exhibits remarkable efficacy, achieving state-of-the-art performance across five publicly available datasets, thus underscoring its prowess in gesture recognition tasks. Furthermore, our approach provides interpretable insights into muscular activation patterns, thereby reaffirming the practical effectiveness of our GCN model. Moreover, we show the effectiveness of our graph-based input structure and GCN-based classifier in maintaining high accuracy even with reduced sensor configurations, suggesting their potential for seamless integration into AI-powered rehabilitation strategies utilizing EMG-based gesture classification systems.

Integrating information from vision and language modalities has sparked interesting applications in the fields of computer vision and natural language processing. Existing methods, though promising in tasks like image captioning and visual question answering, face challenges in understanding real-life issues and offering step-by-step solutions. In particular, they typically limit their scope to solutions with a sequential structure, thus ignoring complex inter-step dependencies. To bridge this gap, we propose a graph-based approach to vision-language problem solving. It leverages a novel integrated attention mechanism that jointly considers the importance of features within each step as well as across multiple steps. Together with a graph neural network method, this attention mechanism can be progressively learned to predict sequential and non-sequential solution graphs depending on the characterization of the problem-solving process. To tightly couple attention with the problem-solving procedure, we further design new learning objectives with attention metrics that quantify this integrated attention, which better aligns visual and language information within steps, and more accurately captures information flow between steps. Experimental results on VisualHow, a comprehensive dataset of varying solution structures, show significant improvements in predicting steps and dependencies, demonstrating the effectiveness of our approach in tackling various vision-language problems.

While exploring visual scenes, humans' scanpaths are driven by their underlying attention processes. Understanding visual scanpaths is essential for various applications. Traditional scanpath models predict the where and when of gaze shifts without providing explanations, creating a gap in understanding the rationale behind fixations. To bridge this gap, we introduce GazeXplain, a novel study of visual scanpath prediction and explanation. This involves annotating natural-language explanations for fixations across eye-tracking datasets and proposing a general model with an attention-language decoder that jointly predicts scanpaths and generates explanations. It integrates a unique semantic alignment mechanism to enhance the consistency between fixations and explanations, alongside a cross-dataset co-training approach for generalization. These novelties present a comprehensive and adaptable solution for explainable human visual scanpath prediction. Extensive experiments on diverse eye-tracking datasets demonstrate the effectiveness of GazeXplain in both scanpath prediction and explanation, offering valuable insights into human visual attention and cognitive processes.

The purpose of this thesis is to propose a 6DoF grasping position measurement strategy for robots based on monocular vision in response to the problems of high cost and poor stability of 6DoF grasping position measurement of target workpieces when industrial robots are grasping metal workpieces under vision system. The strategy adopts a simple structure, low-cost, easy-to-deploy monocular vision system to collect industrial sample data, and through the introduction of virtual reality technology and generative adversarial network, to realize the purpose of data enhancement for industrial metal small sample objects with surface defects and interference from the external environment, to solve the problem of insufficient data due to the difficulty of acquiring data in complex environments, and to achieve the problem of insufficient data due to the later application of monocular vision system as well. The purpose of continuously updating and expanding the sample data can be achieved by applying the monocular vision system in the later stage. At the same time, combining the deep learning-based multi-keypoint object detection technology and 2D-3D affine transformation, the method realizes the accurate measurement of the 6DoF position of the target workpiece. This method can measure the 6DoF position of industrial metal parts only by 2D plane images, and is a new method to estimate the 6DoF pose of a given object from a single RGB image using monocular vision, which possesses the ability to detect the pose of target objects with low-cost, stable, and strong antiinterference, reduces the cost of using robotic vision guidance technology in the industrial field, and provides a new way of thinking for monocular vision to be adopted by the industrial field. Provide a new idea for monocular vision to be adopted by the industrial field. Experiments show that the method has good detection accuracy for industrial small sample objects.

For the problems of insufficient number of samples and lack of sample diversity faced by industrial scenes in robot vision applications, this paper proposes an improved data enhancement strategy for industrial small-sample objects. The method combines a stochastic algorithm with a deep learning-based image generation technique to generate a large number of realistic real-life style target images from a small number of template images, thus effectively improving the recognition ability of target objects in complex backgrounds. In addition, virtual reality technology is introduced to generate virtual artifact images with similar styles to those captured by actual cameras using a virtual engine, further enriching the diversity and coverage of the dataset. This technique also employs an advanced algorithm that incorporates a convolutional network and a self-attention module in a cyclic adversarial generative network (CycleGan) to achieve efficient generation and style migration of industrial object images. This data enhancement strategy enriches the training data of the robot vision system and significantly improves the robot's stable detection ability and data quality of targets in complex environments.