To address insufficient feature representation and eigenbasis misalignment in correspondence calculation for partial 3D shape clusters caused by local missing regions, we propose a correspondence calculation method for partial 3D shape clusters via geometric enhancement and optimized spectral fusion, GeoSpec. First, DiffusionNet is adopted to extract the initial spectral-domain and spatial-domain features on manifolds and to form an initial feature representation. Second, a dynamic neighborhood enhancement mechanism is introduced to strengthen geometric features in missing and boundary regions. Third, an optimized local spectral fusion module is used to fuse the spectral-domain features with the enhanced spatial-domain features, achieving alignment between the spectral basis and the geometric space. Finally, Sinkhorn normalization is applied to regularize the correspondences between each shape and a virtual template, thereby constructing 3D-shape correspondence for partial 3D shape clusters. Qualitative and quantitative comparative experiments are conducted on the SHREC’16 and PFAUST datasets. Compared with the baseline method Wormhole, GeoSpec reduces the average geodesic error by 0.5%, 0.8%, 0.5%, and 1.6% on SHREC’16 CUTS, SHREC’16 HOLES, PFAUST-M, and PFAUST-H, respectively. The experimental results demonstrate that GeoSpec can accurately perform correspondence calculation for partial 3D shape clusters.

To address the challenges of accurately extracting target features from complex scenes in UAV remote sensing imagery and the susceptibility of small objects to being obscured by noise, this paper proposes a lightweight detection algorithm, RE-YOLO, based on YOLOv8n. First, a multi-scale convolutional module named RFCSConv, which integrates channel and spatial attention mechanisms based on Receptive Field Attention Convolution (RFAConv), replaces the original convolution layers. This enhances feature selection and fusion at multiple scales. Second, the Efficient Squeeze-and-Excitation Module (ESEModule) is introduced into the backbone to strengthen feature representation while reducing computational overhead. Lastly, a composite loss function called Win-IoU, combining Wise-IoU (WIoU) and Inner-IoU, is proposed to dynamically adjust gradient contributions based on anchor quality. Experimental results on the VisDrone2019 dataset demonstrate that RE-YOLO achieves 29.7% mAP@0.5 with only 3.2MB of parameters and a real-time speed of 150 FPS. The algorithm also generalizes well across the HRSID and CARPK datasets, achieving 91.8% and 94.3% mAP@0.5 respectively.

In remote sensing imagery, small objects often suffer from inadequate feature representation and are highly susceptible to interference from complex backgrounds, posing significant challenges to detection tasks. Existing detection models are typically characterized by high computational complexity, making efficient deployment on mobile platforms difficult. These issues severely hinder their application and promotion in tasks such as environmental monitoring, rapid military decision-making, and post-disaster emergency response. To address these challenges, this paper proposes a lightweight detection model, termed GAS-YOLO, which leverages contextual information and advanced feature guidance. GAS-YOLO comprises three core modules: a Global Context Extraction Module (GCEM), an Advanced-feature Guided Path Aggregation Network (AG-PAN), and a Shared-Parameter Detection Head (SPD). The synergistic integration of these modules enhances global feature perception, cross-level feature fusion, and global cross-channel correlations, thereby improving the feature representation of small objects, suppressing background interference, and simultaneously reducing model complexity. With only 4.28 GFLOPs and 1.58M parameters, GAS-YOLO achieves detection accuracies of 88.7%, 74.5%, 70.5%, and 45.6% (measured by mAP50) on the USOD, VEDAI (optical and infrared), and AI-TOD datasets, respectively. Compared to the baseline model, these represent improvements of 3.4, 2.7, 3.3, and 2.4 percentage points, outperforming several classical and state-of-the-art algorithms. Experimental results demonstrate that GAS-YOLO effectively improves detection accuracy while significantly reducing complexity by extracting global contextual information and adopting advanced feature guidance strategies.

Using the impressive learning representation capacity of large language models (LLMs), LLM-based video understanding methods have made significant strides recently. However, most existing methods overlook the crucial importance discrepancy of frames, which often include massive low-quality frames, leading to limited performance and inferior inference efficiency, particularly for long-term videos. To this end, this paper proposes a new video understanding method called quality- guided dynamic memory network (QDM-Net). First, we design a memory quality evolution module (MQEM), which dynamically assigns weights to each frame according to contextual relationships between adjacent frames. Second, we devise a high- level quality memory bank updating mechanism (HQMBU), which selectively maintains high-quality frames in the memory bank, avoiding the negative influences of redundant frames and ensuring that the model focuses on the most informative visual cues. Extensive experiments on long-term video understanding benchmarks demonstrate that our QDM-Net consistently outperforms state-of-the-art methods, showcasing its potential in real-world applications. Our code and model will be publicly available.

Existing three-dimensional(3D)point cloud semantic recognition and segmentation algorithms often ignore the relationship between the local feature extraction network and the number of network layers,thus failing to resolve the difficulties associated with expanding the network and capturing advanced semantic information when extracting deeper local features.To address these limitations,an algorithm,namely,semantic recognition and segmentation algorithm of 3D point clouds using a multistage hierarchical fusion residual multilayer perceptron(MLP),is proposed.First,the point clouds are sampled in stages with layered structures to ensure that the network can fully extract feature information at various depths.This involves grouping the sampled points to build local neighborhoods,which enhances the ability of the network to mine local features.Each neighboring domain uses the expandable feature extraction operator of the residual MLP block to extract special information of the cloud.Finally,deep semantic information is integrated with shallow geometric information using interpolation and skip connections.The results reveal that the proposed algorithm achieves a recognition accuracy of 95.1%on the ModelNet40 dataset and a segmentation accuracy of 86.6%on the ShapeNet Part dataset.Thus,this algorithm can effectively extract rich point cloud feature information and offer improved capabilities for 3D cloud semantic recognition and segmentation.

The rapid development of multimodal large language models (MLLMs) has highlighted their potential in video understanding. However, challenges remain in long video tasks, particularly in integrating visual features with prompt texts. Existing methods naively store processed video frames in a long-term memory bank, but neglect simple yet effective cross-modal integration. To address this, we introduce the instruction-aware memory construction (IaMC) model for long-term video understanding. By integrating visual and textual information, our model can obtain cross-modal features with robust understanding capabilities. These features are stored in a text-visual memory bank, enabling efficient long-term aggregation without surpassing LLM context or GPU memory limits. Experiments on the LVU dataset demonstrate state-of-the-art performance in video understanding and question answering, showcasing the IaMC model’s effectiveness and setting a new benchmark for long-term video analysis. The source code and trained models will be released publicly.

This paper proposes an efficient simulation method for modeling of microwave components in multiphysical fields. Firstly, the simulation of electromagnetic model is completed by artificial neural networks (ANNs) finite-difference time-domain (FDTD). The space mesh and time step of the electromagnetic model are enlarged by using the unconditionally stable property of ANN-FDTD to make it consistent with that of the thermal model. Secondly, the thermal modeling and simulation framework leverages Physics-Informed Neural Networks (PINNs) to numerically solve coupled heat conduction equation through deep neural network parameterization. The space mesh and time step of the electromagnetic model are used as the sampling points of PINN. The loss function consists of initial conditions, boundary conditions and physical constraints. This method can effectively avoid the unnecessary calculation of electromagnetic model. The validity of the proposed model is confirmed with one numerical example of a rectangular waveguide filter.

Long video understanding with Large Language Models (LLMs) enables the description of objects that are not explicitly present in the training data. However, continuous changes in known objects and the emergence of new ones require upto-date knowledge of objects and their dynamics for effective understanding of the open world. To alleviate this, we propose an efficient Retrieval-Enhanced Video Understanding method, dubbed REVU, which leverages external knowledge to enhance the performance of open-world learning. First, REVU introduces an extensible external text-object memory with minimal text-visual mapping, involving static and dynamic multimodal information to help LLMs-based models align text and vision features. Second, REVU retrieves object information from external databases and dynamically integrates frame-specific data from videos, enabling effective knowledge aggregation to comprehend the open world. We conducted experiments on multiple benchmark datasets, and our model demonstrates strong adaptability to out-of-domain data without requiring additional fine-tuning or retraining. Experiments on benchmark video understanding datasets reveal that our model achieves state-of-the-art performance and robust generalization.

To address the challenges of low detection accuracy caused by scarce training data for novel classes and class imbalance in remote sensing image datasets for few-shot object detection, we propose an improved fine-tuning-based few-shot object detection model for remote sensing images. First, a multi-branch FPN module is introduced during the fewshot fine-tuning stage. By processing multiple parallel feature branches of varying scales within the network, this module effectively mitigates information loss inherent in traditional architectures. Subsequently, feature maps from different branches are fused to preserve high-resolution details to the greatest extent, significantly enhancing the model's feature extraction capability for novel classes. Second, a balanced loss function module is incorporated into the classification and regression stages of bounding boxes. This loss function dynamically adjusts class weights by accumulating the positive-to-negative gradient ratios for each category, allowing tail classes in remote sensing datasets to receive higher gradient contributions. This design effectively alleviates class imbalance in remote sensing imagery. Under the 10 -shot task configuration, our algorithm achieves mAP improvements of 11.1 % and 3.1 % for novel and base classes, respectively, on the NWPU VHR-10.v2 dataset, and 1.6 % and 2.1 % on the DIOR dataset, compared to the baseline FSCE method.

To address the problem that the correspondence calculation of non-isometric 3D point cloud shape is easily affected by large-scale distortions, which often leads to corresponding distortions, low accuracy, and poor smoothness, a new algorithm of shape correspondence calculation for non-isometric 3D point cloud is proposed, which combines smooth attention with spectral up-sampling refinement. Firstly, a smooth attention mechanism and a smooth perception module are designed using the geometric feature information of the surface on which the points are located to improve the perception ability of the features for non-rigid transformations in large-scale deformation areas. Secondly, the deep functional maps module is combined with smooth regularization constraints to improve the smoothness of the functional maps calculation results. Finally, the final point-by-point mapping result is obtained using a multi-resolution reconstruction method in the spectral up-sampling refinement module. Experimental results show that the proposed algorithm has the smallest geodesic error in the correspondence constructed on the FAUST, SCAPE, and SMAL datasets compared with existing algorithms. It can improve the smoothness and global accuracy of point-by-point mapping for shapes with large-scale deformation.

Electric-hydraulic control system is widely used in subsea oil and gas drilling and production equipment because of its excellent performance. Efficient fault diagnosis of electrical control system helps improve economic benefits and safety of operators. The multi-component characteristics of electro-hydraulic control system make concurrent fault diagnosis a challenging problem. A model-driven method using the Bayesian network and D-S evidence theory is proposed in this paper to diagnose concurrent faults of the electro-hydraulic control system. The electro-hydraulic control system diagnosis problem is divided into multiple fault diagnostic sub-models based on the analysis of system structure and work process. A sub-model is established based on OOBNs for preliminary diagnosis and reasoning. The concurrent fault diagnosis and reasoning model is established based on the D-S evidence theory. The evaluated fault probability and diagnostic belief degree of each fault are combined. Fault identification rules combined with fault probability and diagnostic belief are established. A control system for a subsea blowout preventer is used as a case to demonstrate the proposed method.

Hydraulic control is a control pattern that uses compression fluid as the energy medium and information medium. Hydraulic system is widely used in the control of industrial systems because of its flexibility and reliability. Hydraulic systems have the characteristics of strong fault concealment, significant sensor delay, and a complex signal transmission mechanism. Hence, it is very difficult to identify the faults of systems under the influence of powerful nonlinear time-varying characteristic. The diagnosis of complex hydraulic control system is a problem facing the current research. In order to cope with the challenges caused by limited sensors and concealment of faults, a three-model-driven fault diagnosis method is proposed for complex hydraulic control system. The three hydraulic control system models with regard to energy, fluid and information are proposed and established to explain the work process of the system from different perspectives. The diagnostic model is completed by establishing a fault reasoning model based on the Bayesian network. A redundant control system for subsea blowout preventer is used as a case to demonstrate the proposed method, and the results show that the proposed method has high accuracy.

We provide an innovative methodology for detecting small objects in remote sensing imagery. Our method addresses challenges related to missed and false detections caused by the limited pixel representation of small objects. It integrates super-resolution technology with dynamic feature fusion to enhance detection accuracy. We introduce a cross-stage local feature fusion module to improve feature extraction. In addition, we propose a super-resolution network with soft thresholding to refine small object features, resulting in improving resolution of feature maps while reducing redundancy. Furthermore, we embed a dynamic fusion module based on feature space relationships into a dual-branch network to strengthen the role of the super-resolution branch. Experimental validation on DIOR and NWPU VHR-10 datasets shows mAP improvements to 73.9% and 93.7%, respectively, with FLOPs of 24.89G and 22.33G. Our method outperforms existing approaches regarding accuracy and number of parameters, effectively addressing challenges in small object detection in remote sensing imagery.

针对现有三维点云分类网络采用人工设计费时费力的问题,提出早停可微架构搜索(early-stopping differen-tiable architecture search,ES-DARTS)算法.利用从人工设计网络架构中提取到的先验知识,预定义一个包含高效候选操作的搜索空间,可快速搜索出适用于三维模型分类任务的高性能网络模型;通过追踪网络搜索阶段各候选操作的权重变化,找出跳跃连接操作在双重优化过程中发挥不公平竞争作用的临界点并在此处停止搜索,以保证各候选操作之间的稳定性,解决DARTS算法搜索过程中易出现性能崩溃的问题.提出的算法在ModelNet40 数据集上达到了 93.2％的识别准确率,比当前人工设计的主流网络具有更高的识别准确率.

为满足神经网络中多种位宽数据计算的动态需求,从而提升硬件资源的能效,提出一种位串行乘法器设计—以1 bit的计算逻辑为核心,将多位数据的并行乘操作转化为每个周期进行1位数据乘操作的串行计算方式.为进一步提升硬件资源的利用率,在此基础上提出多通道位串行乘法器阵列同时进行多个数据的并行计算.实验结果显示,在最大支持位宽为8 bit的条件下,单通道位串行乘法器的LUT资源使用量是并行乘法器的41％,LUT资源有效利用率是并行乘法器的1.32倍;当通道数为8时,多通道位串行乘法器阵列的LUT资源使用量是多通道并行乘法器阵列的29％.该结构实现了硬件资源和性能之间的平衡——提高硬件资源的利用率从而提升计算效能.

Non-rigid multi-part shape matching has proven to be essential and challenging in many applications. This paper analyzes the aforementioned problem and proposes a novel multi-part shape matching method to simultaneously compute correspondences between a full shape and its multiple parts undergoing a non-rigid deformation. The main idea is to simultaneously integrate the Hamiltonian eigenvalue equivalence strategy as a part regularization term being fully spectral with the partial functional map. Moreover, we introduce a new upsampling refinement approach based upon ZoomOut in conjunction with the regularized point-wise map recovery algorithm to obtain high-quality partial matches. Our method naturally handles various challenges and noise that commonly occur in real scans, like non-rigid deformations, strong partiality, topological noise, and symmetric ambiguity. Finally, we demonstrate superior qualitative and quantitative results on several datasets. We show that our method produces more accurate, smoother results than other competing methods in realistic scenarios.

针对遥感影像在背景复杂、目标尺度差异性大的场景中检测困难,而现有的 目标检测算法无法同时兼顾检测精度与检测速度等问题,以YOLO v5为基础提出了一种高效的 目标检测网络L-YOLO v5.基于残差空洞卷积对其主干网络进行模型压缩,减少模型的参数量与运算量;为了更好地融合多尺度信息,设计了 一种精细化特征融合网络,从而实现更高效的遥感影像目标检测.在DOTA数据集上的实验结果表明,该算法的mAP和FLOPs分别为77.1％和32.5GB,与YOLO v5相比mAP提高1.2％,FLOPs减少33.7％,在计算复杂度更低的情况下具有更优异的检测效果.

针对现有神经架构搜索算法自动搜索到的网络架构与评估的网络架构之间存在较大差异的问题,提出了基于投票机制的神经架构搜索算法.首先,利用小批量训练数据上测试的训练损失作为性能估计器对候选网络进行采样,将计算资源集中于潜在的性能表现良好的候选网络架构,以解决均匀采样忽略了各网络架构之间重要性程度的问题;其次,对于各节点中候选操作难以选择的问题,利用组稀疏正则化策略对所有候选操作进行排名,以筛选出合适的候选操作,进一步提高Cell结构中路径选择的准确性;最后,将可微架构搜索策略、噪声策略和组稀疏正则化策略加以融合,以加权投票的方法选择出最优的Cell结构,构建出性能优秀的三维模型识别与分类网络架构.在数据集ModelNet40上的实验结果表明,所构建的网络对三维模型的分类准确率达到了93.9％,优于目前的主流算法.本算法有效缩小了搜索和评估阶段网络架构之间的差异,解决了以往神经架构搜索方法中均匀采样所导致的网络训练效率低的问题.

针对现有算法忽略点云数据全局单点特征和局部几何特征的深层关系,导致捕获的局部几何信息缺乏鉴别性且难以有效识别复杂形状的问题,提出基于自注意力特征融合组卷积神经网络的三维点云语义分割算法.首先,设计轻量化网络框架的代理点图卷积提取点云局部几何特征,并加入组卷积操作减少计算量和复杂度,以较少的冗余信息增强特征的丰富性;其次,通过Transformer模块进行不同分支间特征信息的交流,使全局特征和局部几何特征相互补偿,增强特征的完备性;然后,将点云底层语义特征与原始点云融合以扩大局部邻域感受野,获得高级上下文语义信息;最后,将特征输入到分割模块完成细粒度语义分割.实验结果表明,该算法在S3DIS数据集和SemanticKITTI数据集上的分割精度分别达到79.3％和56.6％,能够提取三维点云的关键特征信息,网络参数量较少且具有较高的语义分割鲁棒性.

The calculation of correspondences between 3D models has been widely studied and applied in fields such as autonomous vehicles, virtual reality, and intelligent transportation.However, if the geometric structure and scale of the 3D models compared differ substantially, features extracted by low-level geometric information descriptors are insufficient, as is the accuracy of the results of such correspondence calculations.Therefore, this paper proposes a method to calculate the correspondence of 3D models by introducing a priori knowledge.A deep learning network is used to simulate human prior knowledge to encode geometric similarities between the parts of two models.The proposed method solves the problem that low-level geometric information cannot be used to calculate correspondence relationships between models when their parts differ significantly.A multi-view convolutional neural network is used to pre-segment and mark the corresponding views of each part of the model, to implicitly calculate a data-driven descriptor according to the similarity between corresponding surface points, and finally to calculate the output correspondence between the two 3D models under the guidance of the data-driven descriptor.The experimental results show that the proposed method was able to improve the accuracy of the results of calculations of correspondence relationships between 3D models compared with a calculation method based on a priori knowledge, and it can effectively reduce the geodetic error.