最新刊期
卷 53 ， 期 2 ， 2025
- 摘要：Foundation models deployed in dynamic edge environment encounter continuously evolving input data distributions, requiring retraining them to maintain high accuracy. However, existing retraining techniques can only train fixed compressed models within the constraints of device resources and retraining windows, thus considerably lowering accuracies due to these small models’ limited generalization ability. For such an issue, this paper proposes BlockTrainer, an edge-cloud collaborative retraining approach of foundation models at the block granularity. BlockTrainer first introduces a model retraining scaling law to evaluate the accuracy contributions of different blocks in a foundation model according to its latest input data at edge. Based on this evaluation, it generates the optimal retraining solution under resource constraints, and dynamically converts the most accuracy-relevant parts of the model into retrainable small models at edge, thereby constructing a collaborative training system between large and small models. Comparative experiments on real edge-cloud platforms show that BlockTrainer improves the retraining accuracy of foundation models by 81.24% using the same resource consumptions, and supports retraining a model of up to 33 billion parameters.关键词：foundation model;dynamic environment at edge;model retraining;scaling law;edge-cloud collaborative retraining of large and small models85|71|0更新时间：2025-12-08
- 摘要：Artificial intelligence technology is widely used in marine meteorological forecasting, and increasingly relies on the internet of things to obtain massive sensory data of multiple modes in the marine environment. Aiming at the problem that the amount of data obtained and the transmission speed are not enough to support the accurate forecast of the model, a low-latency artificial intelligence of things (AIoT) construction scheme for marine meteorological forecasting is proposed. Firstly, an AIoT marine meteorological forecast fusion architecture is designed. This architecture not only fully adapts to artificial intelligence technology and internet of things technology, but also realizes the efficient collection, processing and analysis of marine meteorological data, providing an effective and flexible underlying structure for marine meteorological forecasting. Secondly, the collaborative networking method of heterogeneous ocean sensing devices is optimized. By optimizing the multi-layer coupling network topology, efficient interconnection of heterogeneous ocean sensing devices was achieved, ensuring the comprehensiveness and accuracy of data collection. Finally, a low-latency routing algorithm for ocean sensing networks is proposed. This algorithm reduces the transmission delay of information from sensing devices to data centers through intelligent path selection and data transmission optimization, ensuring the rapid update of forecast data. Experimental verification shows that the proposed scheme fully utilizes the advantages of the AIoT and solves the problems of difficult data acquisition and long processing delay in marine meteorological forecasting. The mean delay of the proposed scheme is reduced by 37% and the median delay is reduced by 38%, providing strong support for real-time and accurate marine meteorological forecasting.关键词：artificial intelligence;Internet of Things;marine meteorological forecasting;network optimization;routing strategies50|31|0更新时间：2025-12-08
- 摘要：With the widespread adoption of internet of things (IoT) and smart devices, the volume of data generated at the edge has far exceeded the computational and storage capabilities of edge nodes. This creates an urgent need for cloud-edge collaborative processing to meet the real-time analysis demands of large-scale data. With the decoupling of computation, memory, and storage, the shared-cahe architecture has become a critical solution for addressing the processing requirements of massive edge data. However, there are still several issuses remained in shared-cache architecture. First, in transactional processing scenarios, when hotspot cached data frequently migrates between nodes, the log persistence mechanisms of existing databases will generate a large number of log write operations, thereby impacting system performance. Secondly, the existing cache write-invalidation mechanism can lead to frequent eviction of some hotspot cached data, causing slower transactions to fail in retrieving target data from the shared cache in time. This could trigger a large number of cache reloads, resulting in system performance degradation. To address these issues, this paper proposes a dependency-table-based delayed log flushing mechanism. By consolidating multiple log write operations and deferring them until the log buffer is full or a transaction is committed, the mechanism reduces the frequency of log flushing and the overhead of disk writes. In addition, this paper also introduces a cache delayed invalidation mechanism that incorporates asynchronous replay of invalidation messages, page visibility determination, and an optimized cache replacement. This approach effectively extends the service time of cached data, improving cache hit rates and overall system performance. Based on these mechanisms, this paper implements a high-performance shared-cache database system called EBASE-T. Experimental results show that, compared to its pre-optimized version, EBASE-T achieves a 19.5% increase in throughput and a 13.1% reduction in latency. In TPC-C (online transaction processing system benchmarks) tests, EBASE-T demonstrates significant performance advantages over most shared-cache database systems.关键词：database;distributed shared cache;transaction processing;log flushing;cache invalidation41|21|0更新时间：2025-12-08
- 摘要：Vehicle edge computing combines mobile edge computing and the internet of vehicles(IoV) to offload the vehicle computing tasks from the cloud servers to edge servers, which effectively reduces the response time of IoV services. However, the irregular spatiotemporal distributions of traffic flows in vehicle networking will lead to the imbalance of computing load on the edge servers, which impacts real-time responsiveness of vehicle networking services. To address this issue, this paper proposes an efficient task offloading strategy based on traffic prediction in the vehicle edge computing. Specifically, a chebyshev graph weighted network (ChebWN) is designed to forecast traffic flow by fully leveraging connectivity and distance information between road segments. Next, a deep reinforcement learning-based binary task offloading algorithm (DBOA) is designed, which divides the binary task offloading decision process into two stages. Initially, a deep reinforcement learning approach is employed to derive the offloading strategies. Subsequently, a one-dimensional bi-end search algorithm is utilized to determine the time slot allocation scheme that maximizes the overall computation rate, thereby reducing the complexity of the decision-making process. Finally, a large number of comparative experiments demonstrate the accuracy of ChebWN in predicting traffic flow and the superiority of DBOA in improving the response speed of vehicle services.关键词：mobile edge computing;deep reinforcement learning;internet of vehicles;graph neural network;task offloading87|50|1更新时间：2025-12-08
- 摘要：Erasure coding and near-data processing are two cornerstones for building efficient cloud-edge-end collaborative data management systems. The former ensures system availability by adding coding redundancy to the original data, while the latter avoids significant network transmission overhead by processing data at the storage end. Cloud-edge-end collaborative data management systems typically adopt mature distributed storage systems as the underlying storage engine. However, the erasure coding implementation in mainstream distributed storage systems can not efficiently support near-data processing. This paper proposes an erasure coding architecture that supports near-data processing. Its basic principle is to reorganize the data to be encoded to ensure that semantically related data is stored in the same storage device, thereby avoiding cross-node data transmission during near-data processing. The scheme has been implemented on the distributed storage system Ceph, and the read and write performance under typical scenarios are tested. The experimental results show that the performance of reading objects in the near-data processing scenario and the conventional data reading scenario are improved by about 59.4% and 10% respectively, while the object writing performance remains consistent with the original version.关键词：erasure coding;distributed storage;Ceph;near-data processing;cloud-edge-end collaborative data management26|37|0更新时间：2025-12-08
- 摘要：With the advancement of modern communication and information technology, intelligent transportation systems(ITS) have emerged as a prominent area of research. The vehicular ad hoc network(VANET), serving as its pivotal technology, plays a crucial role in facilitating real-time road information sharing and inter-vehicle communication. However, existing clustering algorithms for VANET are plagued by issues such as low stability and high overhead. To address these challenges, this paper proposes a VANET clustering algorithm that leverages end-cloud collaboration. In the end-cloud collaboration phase, vehicles upload their feature data to the cloud via road side units(RSU), where the cloud performs dynamic stability classification based on changes in vehicle features. Nodes exhibiting stable behavior demonstrate higher reliability and longer connection durations. In the end-to-end coordination phase, factors including relative node mobility and cluster coverage are taken into account during cluster-head election to streamline the process while enhancing cluster stability. Furthermore, this paper introduces a neighbor discovery and update mechanism aimed at restricting HELLO message forwarding operations to reduce overhead and optimize resource utilization. Experimental results demonstrate that the proposed algorithm surpasses baseline algorithms across key performance metrics such as cluster stability, quantity of clusters formed, and clustering costs—highlighting its potential applicability in real-world traffic scenarios.关键词：vehicular ad hoc network;multi-hop clustering;stable node;cluster head election;end-cloud collaboration45|27|0更新时间：2025-12-08
Cloud-Edge-Device Collaborative Data Management, Analysis and System
- 摘要：With the vigorous development of new technologies and services such as artificial intelligence, big data, cloud computing, etc., the communication paradigm between Internet applications, which plays a supporting role, has evolved from the traditional “request-get” one-stage to the complex “request-compute-get” two-stage. The new communication paradigm requires the network not only to provide the data transmission channels but also to assume the role of opening up computation in the form of public services on the Internet, thus deepening the integration of computing and network and promoting the development of the digital economy. However, in the network architecture represented by the current TCP/IP architecture, there are some drawbacks in the core mechanisms, including resolution and discovery, network awareness, and route computation, which make it difficult to become a foundation for new applications oriented to the convergence of communication and computing. In this paper, we propose a new implementation architecture named service aware network (SAN), which systematically changes the existing design of host interconnection to provide comprehensive, open, and guaranteed endogenous service interconnection support for new scenarios such as computing network convergence. In SAN, a semantic service identifier enables user terminals to initiate location-independent connections based on service types. Computing and network demand awareness provides a proactive method for the network to obtain the requirements of requests. “Two-stage two-layer” routing not only realizes the routing considered two-dimensional computing and bandwidth resources but also guarantees the feasibility and validity of SAN deployment under different degrees of network openness. To quickly implement the concept of SAN, the SAN architecture avoids major transformation of the existing network through incremental deployment. It realizes new mechanisms without affecting the functions of the existing network and promotes the smooth evolution of the network from host interconnection to service interconnection. Experiments show that the SAN routing mechanism reduces the upper bound of service completion time by 35.4% and 17.5%, respectively, compared with the benchmark routing algorithms, and achieves more balanced and efficient computing and network resource usage.关键词：service aware network;network architecture;service identifier;network and computing demand awareness;routing44|32|0更新时间：2025-12-08
- 摘要：Aiming at the problem that traditional frequency hopping network station sorting technology is ineffective under low signal-to-noise ratio conditions and has poor real-time detection performance, this paper proposes a shortwave frequency hop-ping signal sorting algorithm based on the improved YOLOv8 (You Only Look Once version 8). First, the short-time Fourier transform is performed on the received aliasing signal to generate a grayscale time-frequency image as the input of the YOLOv8 network model. Secondly, in view of the impact of frequency collisions between aliasing signals such as sweep frequency signals, fixed frequency signals and frequency hopping signals on detection accuracy, the Deformable Convolutional Net-works v2 is introduced in the C2f layer to improve the generalization ability of network feature extraction. Thirdly, the Simam attention mechanism is added to the backbone layer to solve the problem that background noise is easily confused with frequency hopping signals and affects detection accuracy under low signal-to-noise ratio. Finally, the convolutional kernel of Detect module is replaced by Partial Convolution kernel, which reduces the computational complexity of the network by 32.18% without the accuracy loss of mAP@0.5 exceeding 0.37%, and improve the inference speed of the network model. Experimental results show that the improved YOLOv8 algorithm proposed in this paper has a separation rate of 97.68% at -5 dB signal-to-noise ratio, and the model has fast convergence and strong robustness.关键词：frequency hopping signal sorting;YOLOv8;DCNv2;SimAM mechanism;PConv71|37|0更新时间：2025-12-08
- 摘要：Most of the traditional tunable terahertz broadband absorbers are characterized by complex structure and single function, which make their practical application scenarios limited. To address this problem, this paper proposes a bifunctional THz metamaterial absorber based on the thermal phase transition property of vanadium dioxide (VO2) and the pump light sensitivity property of photosensitive silicon (Si). By adjusting the conductivity of VO2 and Si, not only the dynamic switching between broadband absorption and electromagnetically induced transparency (EIT) can be realized, but also the amplitude values of the absorption rate and the resonance peak of EIT can be dynamically adjusted. When σ(VO2) = 2×105 S/m and σ(Si) = 10 S/m, the absorber exhibits broadband absorption in the range of 1.24~3.65 THz with an average absorbance more than 94%. When σ(VO2) = 20 S/m and σ(Si) = 1×106 S/m, the transmittance transmission curve of the absorber exhibits the EIT effect, which creates a resonant transparent window with an amplitude greater than 90% in the range of 1.50~2.50 THz. In addition, the physical mechanism of the absorber was further analyzed using impedance matching and equivalent circuit theory. Subsequently, the refractive index sensing characteristics of the EIT resonant peaks were analyzed, and the sensing sensitivities of f1 and f2 were 224 GHz/RIU and 310 GHz/RIU with Q values of 3.82 and 18.5, respectively, which provided good sensing performance. Finally, the slow light effect of the EIT window is analyzed, and the results show that the device has good slow light effect. With simple structure, wide operating bandwidth, large tunable range and dual-function switchable, the absorber proposed in this paper has potential future applications in the fields of optical devices, electromagnetic stealth and sensing detection.关键词：terahertz;metamaterial absorber;broadband absorption;electromagnetic induced transparency29|28|0更新时间：2025-12-08
- 摘要：The existing passive localization methods assume that the moving sensor positions remain unchanged (or approximately unchanged) during signal propagation. This assumption is reasonable in the free space of wireless radio propagation. However, in environments where slow signals such as acoustic waves or seismic waves are used for localization, the relative motion between the sensors and the unknown targets results in a non-negligible ratio of the movement distance to the signal propagation distance, using existing passive localization models may lead to substantial performance loss. To address this issue, this paper take the underwater acoustic localization scenario as an example to investigate the passive localization algorithm in the presence of the Sensor Motion Effect (SME) and the unknown underwater acoustic signal propagation speed. We first establish the observation model for underwater passive localization with SME, analyze the performance loss due to neglecting SME in existing localization models, and derive the Cramer-Rao Lower Bound (CRLB) for the new localization scenario. Subsequently, the closed-form solution is developed for determining the source position and the unknown underwater acoustic signal propagation speed. The effectiveness and superiority of the proposed method are validated through theoretical analysis and numerical simulations.关键词：passive localization;sensor motion effect;unknown underwater acoustic signal propagation speed;Cramer-Rao Lower Bound;closed-form solution39|28|0更新时间：2025-12-08
- 摘要：During the rapid evolution of post-quantum cryptography, considering the needs for flexibility and efficiency, we proposed a parallel reconfigurable sampling accelerator for various lattice-based post-quantum cryptographic algorithms. We analyzed seven sampling processes involved in lattice-based post-quantum cryptography and proposed seven efficient parallel implementation models for these samplings, respectively, based on mathematical derivations. Then we extracted four common operational logics from these models. Using these four common operational logics as the core, we introduced data rearrangement to limit the effective bit width of operation data, which improved the acceptance rate of rejection sampling and eliminates the complex modular reduction operations in finite field operations. Then we proposed a high energy-efficient reconfigurable parallel sampling algorithm. To enhance the hardware implementation efficiency of the sampling algorithm, we adopted the butterfly transform network to complete the parallel splitting, merging, and lookup of data with any effective bit width within a single clock cycle, efficiently realizing the parallelization of the algorithm’s pre- and post-processing, and constructed a parameterized parallel reconfigurable sampling accelerator architecture model. Aiming for high energy efficiency, combined with logic synthesis experimental results, we determined the optimal parallel degree parameters of the architecture model and proposed a parallel reconfigurable sampling accelerator with a data bandwidth of 1 024 bits. Experimental results showed that, using a 40 nm CMOS process library, and performing post-simulation under the ss, 125 ℃ process corner conditions, the circuit's highest operating frequency can reach 667 MHz, with an average power consumption of 0.54W. Completing a 256-point uniform sampling requires 6 ns, completing a 256-point rejection sampling with a rejection value less than 216 on average only takes 22.5 ns, completing a 256-point binary sampling within 8 bits requires 18 ns, completing a 509-point simple ternary sampling requires 36 ns, completing a 701-point non-negative correlated ternary sampling requires 124.5 ns, completing a 509-point fixed-weight ternary sampling requires 11.18
- 摘要：To solve the compatibility problem of multi-band on a single chip in the RF front-end of the receiver, this paper proposes a new bandwidth-reconfigurable low noise amplifier (LNA) structure for UWB applications. This LNA is based on switchable reconfigurable design methods, embedding the switchable design in the load of the cascaded LNA circuit. The design achieves switching of in-band input impedance matching and gain curves for different UWB operating modes by controlling the position of low-frequency impedance resonance point and corresponding gain pole through the reconfigurable design of the load inductance of the resistive parallel negative feedback structure. Compared with the design methods of introducing switches in the input/output matching path, placing switches at the load optimizes gain and noise performance without affecting impedance matching. The resistors and inductors in the traditional inductive peaking technique are adjustable to consider gain flatness within different operating bandwidths. Based on SMIC 28 nm CMOS technology, the simulation results of electromagnetic modeling demonstrate that the LNA operates in three modes: 3.1~10.6 GHz, 6~10.6 GHz, and 3.1~5 GHz, with in-band voltage gain (S21) above 16.59 dB and minimum noise figure below 3 dB. Under 0.8 V power supply voltage, all three modes exhibit input and output matching (S11, S22) below -10 dB, with a static power consumption of only 9.03 mW; after introducing MOS switches, the noise figure degradation of the LNA in all three bandwidths is less than 0.2 dB.关键词：low noise amplifier;ultra wide band;adjustable bandwidth;reconfigurable;switching;inductive peaking65|86|0更新时间：2025-12-08
- 摘要：To address the significant time overhead and free-rider effect in most GNN-based community search methods, this paper proposes an efficient community search based on graph combinatorial optimization (CS-ROMF). CS-ROMF designs a GNN-based community locator to quickly pinpoint potential communities of the query nodes, thereby reducing time overhead. On this basis, CS-ROMF further designs an RL-based community optimizer to adjust the structure of candidate communities, mitigating the free-rider effect. Experiments conducted on five real-world datasets with true communities demonstrate that CS-ROMF outperforms advanced community search methods across all evaluation metrics. Specifically, compared to the best baseline model, CS-ROMF achieves maximum improvements of 14.99%, 20.67%, and 21.37% in F1 score, Jaccard score, and NMI, respectively. Additionally, CS-ROMF can significantly improve search efficiency, running up to 10 times faster than the baseline model based on GNN.关键词：community search;graph neural network;graph combinatorial optimization;matching strategy;reinforcement learning;community detection35|48|0更新时间：2025-12-08
- 摘要：Transferring product to the entity which is not trusted by the administrator may lead to some problems, such as product counterfeiting, smuggling, product loss, and privacy leaking, etc. Therefore, in this paper, a product transfer scheme named BPOTS in RFID-enabled supply chain based on blockchain is proposed. Firstly, this paper proposes a secret value sharing and verification algorithm based on Chinese remainder theorem and Pedersen commitment to achieve the transfer of products between the designated new owner sets. And in order to improve system efficiency, we propose a method for the transfer of products in batches based on the homomorphism of Pedersen commitment. Secondly, to balance the transparency and privacy of the supply chain, this paper proposes a pseudo ID generation algorithm based on symmetric encryption. Thirdly, security analysis and performance evaluation are conducted on the BPOTS scheme. The result shows that BPOTS strikes a balance between the transparency and privacy of the supply chain effectively and improves the efficiency of transferring product for about 12 times compared with the existing product ownership transfer schemes. Finally, the BPOTS scheme is implemented on ChainMaker platform and made available as open-source on Github. The testing result indicates that the efficiency of transferring product in BPOTS scheme is about 70.4% higher than that of transferring products in series. Moreover, BPOTS scheme reduces the costs of supply chain nodes effectively.关键词：blockchain;RFID-enabled supply chain;ownership transfer;Chinese remainder theorem;Pedersen commitment33|26|0更新时间：2025-12-08
- 摘要：Most current blockchain systems can hardly concurrently satisfy requirements of privacy protection and transaction data auditing. Bitcoin adopts the method of unspent transaction output (UTXO) to ensure that users can quickly query the source as well as fund destinations of each transaction, in order to avoid double spending threats. However, the users’ behaviors, deemed to be privacy, maybe traced by adversaries, since transactions with addresses are stored in the ledger publicly. Even though encryption-based solutions are widely adopted, it often causes restrictions to transaction verifications and auditing. In this paper, we propose an auditable privacy-preserving confidential transaction scheme, which uses Pederson commitment to realize the public verifiability of the transaction rationality without disclosing the specific amount of the transaction. Our scheme supports the initiator of the transaction to initiate the transaction independently without permissions from the receiver, which saves the communication cost comparing with other confidential transaction schemes. By introducing the trapdoor mechanism, the identity of the transaction initiator cannot be recognized by other users outside the ledger and the supervisor, so as to protect users’ privacy. It has realized a variety of audit functions, and different audit methods have been developed according to regulators and private auditors. This paper presents a new range proof method, which has advantages over Prcash when applied to large numbers. The generation time of range proof for 512 bit large numbers is shortened 29.78%, and the generation time of range proof for 1 024 bit large numbers is reduced 56.86%.关键词：auditable;zero-knowledge proof;Pederson commitments;homomorphic encryption;range proof58|33|0更新时间：2025-12-08
- 摘要：IntervalTree is a search tree used for maintaining dynamic sets, specifically designed for efficient storage and retrieval of interval collections. The current implementation of IntervalTree involves modeling and verification in Isabelle/HOL, where interval information is expanded upon a binary search tree. However, the time complexity of the basic operations supported by the IntervalTree structure is relatively high. To address this issue, this paper proposes the IntervalTree+ structure, augmenting nodes of the IntervalTree with additional color information to ensure tree balance. As compared to the original IntervalTree structure, the worst-case time complexity for operations such as insertion and deletion is improved from O(n) to O(log n) in the IntervalTree+ implementation. Subsequently, functional modeling of the IntervalTree+ structure and its operations is performed using the Isabelle theorem prover. Mechanical verification of invariants is conducted to ensure the correctness and reliability of IntervalTree+ structure operations. Additionally, for the first time, a generic verification specification for region matching algorithms is proposed to address correctness verification issues across a series of such algorithms. The proposed IntervalTree+ structure has been rigorously verified through formal mechanization. Compared to IntervalTree structure, its worst-case time complexity is optimized from O(n) to O(log n). This optimization makes it applicable to algorithmic enhancements in areas such as region matching, visual logging, and model evaluation.关键词：IntervalTree;IntervalTree+;functional modeling;mechanized verification;region matching algorithm;Isabelle theorem prover19|26|0更新时间：2025-12-08
- 摘要：As the recording density of magnetic storage increases, the recording bit spacing decreases and the magnetization transition noise increases significantly, which greatly affects the quality of the readback signal. To mitigate the interference of magnetization transition noise problem among recording patterns in ultra-high density magnetic storage systems, the maximum transition run (MTR) constraint code MTR(
- 摘要：In recent years, the contrastive language-image pre-training (CLIP)-based method takes the text information as the weight of the classifier through the joint representation of image and text and shows excellent performance in the general image recognition task. However, the existing methods only construct text prompts of categories, such as context optimization (CoOp) and conditional context optimization (CoCoOp), without considering the importance of image content semantic information and categories, which limits the model’s understanding and discrimination of image categories. To solve the above problems, this article proposes a new method based on CLIP: discriminative category prompt learning based on image content understanding (DCPL), which uses rich content features in images to learn text prompts and introduces manual templates to improve the discrimination of text prompts on categories. Specifically, DCPL includes a prompt generation module and a text supervision module: The prompt generation module takes image features and initialized query vectors as inputs, and makes the output text prompt contain sufficient image semantic information through the self/cross-attention mechanism; The text supervision module uses the fixed category prompt template as the supervision to inject category information into the category level and logits level for the learnable text prompt, increasing the importance of categories. Finally, the average accuracy of 16 shots of DCPL on 11 public classified datasets, such as ImageNet, Caltech101, Oxford pets, etc., is 81.84%, The average accuracy has increased by 0.98 percentage points compared with that of the previous optimal method, Cross-Modal.关键词：visual-language model;image recognition;prompt tuning;attention mechanism;text supervision;adapter tuning;Transformer25|47|0更新时间：2025-12-08
- 摘要：Atrial fibrillation (AF) is a common arrhythmia often associated with cardiovascular diseases such as stroke and heart failure. Although numerous researchers have made substantial progress in AF detection using deep learning methods in recent years, most of these methods require extensive computational resources. Moreover, the clinical application of these models is challenging due to the black-box nature of deep learning models. Therefore, this paper proposes a lightweight AF detection model based on feature fusion and conducts an interpretability study. The model comprises an ECG (ElectroCardioGram) backbone network and an RRI (R-R Interval) branch. The ECG backbone network uses depthwise separable convolutions along with a few standard convolutions to extract deep morphological features of the ECG signals, while the RRI branch employs multi-scale convolutions to extract deep rhythm features of the RRI. The network learns robust feature representations by fusing morphological features and rhythm features to detect AF accurately. As to interpretability analysis, Grad-CAM++ is utilized to visualize the contribution of different features to the classification results. In this paper, the training and dataset internal tests are conducted in the LTAFDB and achieved an accuracy of 97.99%. In order to validate the generalization performance of the model, external testing experiments are conducted using the AFDB and the CPSC2021, achieving an accuracy of 95.17% and 93.81%, respectively. Experimental results demonstrate that the proposed method is lightweight, stable, and accurate, and the incorporation of interpretable deep-learning techniques suggests that the proposed method holds significant potential for the clinical diagnosis of AF.关键词：electrocardiogram(ECG);atrial fibrillation;lightweight neural network;visualization technologies;feature fusion56|25|0更新时间：2025-12-08
- 摘要：Images obtained in low light environments often have low brightness, low contrast, and uneven lighting, resulting in weakened and blurred image features that are difficult to extract. At the same time, there is also a large amount of noise information in the limited extracted features, making it difficult to detect and recognize objects. Therefore, there are very few existing low light object detection results. This paper proposes a low illumination object detection method based on the Illumination Correction and Feature Interaction Enhancement (ICFIE-YOLO) network to address the difficulties in extracting features from low illumination objects and the large noise in the feature space. This method first utilizes the proposed ICFIE-YOLO internal Multi Scale Illumination Correction Network (MSICN) to correct low illumination images, highlighting the blurry features of objects hidden in the image’s background, and enabling the feature extraction module to better extract object features. Secondly, to fully utilize effective feature information and filter out noise interference in feature maps, a Feature Interacted Enhancement (FIE) detection head is proposed. Through feature attention interaction, feature enhancement is achieved, establishing spatial and semantic correlations between features in different regions of low illumination images, thereby suppressing the interference of noise on effective features and achieving feature enhancement. Finally, on the basis of enhancing features and removing noise, an improved detection head is used to achieve high-precision object detection. Experiments on the ExDark and DarkFace datasets show that the proposed Method improves mAP by over 2.1 percentage points compared to mainstream object detection models, increases recall by over 4.2 percentage points compared to existing low light object detection Methods, and improves recall by 2.6 percentage points compared to baseline models. The proposed Method has good generalization performance.关键词：object detection;Low illumination;Light correction;Feature denoising;Feature enhancement74|31|0更新时间：2025-12-08
- 摘要：Enhancing the quality of underwater images is crucial for advancements in the fields of underwater exploration and underwater rescue. Existing underwater image enhancement methods typically rely on paired underwater images and reference images for training. However, obtaining corresponding reference images for underwater images is challenging in practice. In contrast, acquiring high-quality unpaired underwater images or images captured on land are relatively more straightforward. Furthermore, existing techniques for underwater image enhancement often struggle to address a variety of distortion types simultaneously. To avoid the reliance on paired training data, reduce the difficulty of acquiring training data, and effectively handle diverse types of underwater image distortions, in this paper, we propose a novel unpaired underwater image enhancement method based on the frequency-decomposed generative adversarial network (FD-GAN). We design a dual-branch generator based on high and low frequencies to reconstruct high-quality underwater images. Specifically, feature-level wavelet transform is introduced to separate the features into low-frequency and high-frequency parts. Then the separated features are processed by a cycle-consistent generative adversarial network, so as to simultaneously enhance the color and luminance in the low-frequency component and details in the high-frequency part. More specific, the low-frequency branch employs an encoder-decoder structure with a low-frequency attention mechanism to enhance the color and brightness of the image. The high-frequency branch utilizes parallel high-frequency attention mechanisms to enhance various high-frequency components, thereby achieving the restoration of image details. Experimental results on multiple datasets show that the proposed method trained with unpaired high-quality underwater images or unpaired high-quality underwater images and on-land images, can effectively generate high-quality underwater enhanced images and the proposed method is superior to the state-of-the-art underwater image enhancement methods in terms of effectiveness and generalization.关键词：underwater image enhancement;generative adversarial networks;wavelet transform;attention mechanism;dual-branch generator based on high and low frequencies39|33|0更新时间：2025-12-08
- 摘要：Sleep is the significant physiological process to keep healthy. Sleep stage classification based on polysomnography (PSG) is the fundamental evidence to diagnose sleep disorders and assess sleep quality. Manual sleep staging method has some typical problems when handling the large-scale PSG data, such as time-consuming and low-efficiency. The automatic sleep staging method that utilizing deep learning models to effectively learn PSG representations shows extensive researching prospects. Most existing models cannot fully consider the epoch-level waveform information, channel-wise correlations, sequence-level sleep transitions. This paper proposes a transformer-based hierarchical sleep staging model (HierFormer), which employs transformer encoder to extract valid epoch-level waveform features, channel-wise correlation features, sequence-level transition features. Meanwhile, it adopts attention mechanism to improve the model interpretability of signal properties from three views, including epoch-level, channel-wise, and sequence-level views. Experimental results on the sleep-european data format (sleep-EDF) database expanded dataset show that the proposed model achieves better sleep staging performance with less parameters compared with various baseline models. The overall accuracy, macro-averaging precision, macro-averaging recall, macro-averaging F1-score, and Cohen’s-kappa coefficient achieve 0.807, 0.784, 0.735, 0.750, and 0.721, respectively. According to the performance comparisons of different feature encoding methods from three views and the visualization of attention weights, this paper further demonstrates the satisfied encoding ability and interpretability of proposed model. This study aims to provide innovative deep learning approaches and technologies for the research of sleep staging applications, thus assisting sleep experts to improve the efficiency of sleep disorder diagnosis and treatment.关键词：polysomnography (PSG);automatic sleep stage classification;deep neural network;transformer architecture;attention mechanism;model interpretability29|35|0更新时间：2025-12-08
- 摘要：With the explosive growth of image data in the field of e-commerce, target image retrieval has become a challenging work in information retrieval research. The existing traditional image retrieval models only rely on a single text description or similar image, which is difficult to accurately capture the user’s retrieval intention, resulting in unsatisfactory retrieval results. In order to solve this problem, this paper proposes a combined retrieval method that fuses image and text features. Swin Transformer (SwinT) is used to extract the multi-layer features of the reference image, and the image and text features are fused at multiple levels, so that the text features can modify the reference image features at multi-level and fine-grained, and get closer to the target image features. Then, the modified image features and the target image features are embedded in a space for similarity measurement, and the batch-based classification loss is used to optimize the retrieval performance. Experimental results on Fashion200k, MIT-States and CSS datasets show that the proposed method improves the performance by 5 percentage points on average compared with the existing mainstream methods.关键词：combined image and text retrieval;image features;text features;features fusion804|195|0更新时间：2025-12-08
- 摘要：Restoring the weight distribution, activation distribution, and gradient to the original full precision network data as much as possible can greatly improve the inference ability of the binary network. However, existing methods directly apply the restoration operation in forward propagation to binary data, and the gradient approximation functions for backpropagation are fixed or manually determined, resulting in the need for improvement in the restoration efficiency of binary networks. To address this problem, the efficient restoration method is investigated for binary neural networks. Firstly, a distribution recovery method for maximizing information entropy is proposed. By shifting the original full precision weight mean and scaling the modulus, the quantized binary weight directly has the characteristic of maximum distribution restoration. At the same time, a simple statistical translation and scaling factor is used to greatly improve the restoration efficiency of weight and activation. Furthermore, it is proposed a gradient function based on adaptive distribution approximation, which dynamically determines the update range of the current gradient in the P-percentile according to the actual distribution of the current full precision data. It adaptively changes the shape of the approximation function to efficiently update the gradient during the training process, thereby improving the convergence ability of the model. On the premise of ensuring the improvement of execution efficiency, theoretical analysis has confirmed that the method proposed in this paper can achieve maximum restoration of binary data. Compared with the existing advanced binary network models, the experimental results of our method show excellent performance, with a 60% and 67% reduction in computational time for the distribution restoration operation quantization of ResNet-18 and ResNet-20, respectively. An accuracy of 93.0% is achieved for VGG-Small binary quantization on the CIFAR-10 dataset, and 61.9% is achieved for ResNet-18 binary quantization on the ImageNet dataset, both of which are the best performance of the current binary neural network. The relevant code is available inhttps://github.com/sjmp525/IA/tree/ER-BNN.关键词：Binary neural network;Information restoration;Maximum information entropy;Adaptive gradient37|37|0更新时间：2025-12-08
- 摘要：In this paper, we analyze the fusion law of scene weak change region and type change region, the difference of physical model and the complementarity of effect from spatial and temporal dimensions, and propose a shared super-resolution dual-branch (Shared Super-Resolution Dual-Branch, SSRDB) remote sensing image spatio-temporal fusion algorithm. The algorithm has the following three characteristics: (1) A complementary network framework is constructed. Although the framework is an end-to-end deep learning model, each module has its own physical meaning and task. By adding intermediate supervision, the super-resolution modeling of spatial dimension, the change prediction modeling of time dimension and the fusion modeling of the two advantages are realized respectively. (2) The mathematical representation of the change prediction is deduced, and a nonlinear compensation module is used to make the two branches share the super-resolution module. Under the dual strategy of sharing super-resolution module and recursive multiplexing super-resolution unit, the network parameters are significantly reduced. (3) The recursive super-resolution module uses fixed 2-magnification super-resolution units to gradually enhance features and reconstruct images under effective supervision and reference, which can effectively improve the precision of super-resolution, and flexibly adapt to spatio-temporal fusion tasks with different magnification differences by adjusting the number of super-resolution units. The SSRDB algorithm shows excellent fusion effect in spatial and spectral characteristics and change regions. The three quantitative evaluation indexes of RMSE (Root Mean Squared Error)、SAM (Spectral Angle Mapper) and SSIM (Structural Similarity) show that it is superior to the sub-optimal method on the CIA (Coleambally lrrigation Area) dataset by 7.067%, 2.065% and 0.563%, respectively. On the LGC (Lower Gwydir Catchment) dataset, it is superior to the sub-optimal method by 5.319%, 5.490% and 1.455%, respectively. On the Nanjing dataset, it is superior to the suboptimal method by 6.486%, 16.290% and 0.481%, respectively.关键词：remote sensing image;spatiotemporal fusion;dual-branch;image super-resolution;convolutional neural network33|45|0更新时间：2025-12-08
- 摘要：Freezing rain is recognized as a common hazardous weather condition that often results in ice accumulation on high-voltage transmission lines. When the ice reaches a certain thickness, the shedding of ice cover can trigger flashover between grounded wires, leading to widespread interruptions in urban power supply. In regions prone to frequent freezing rain, power lines experiencing ice shedding oscillations exhibit significant characteristics of wire displacement and tension variation compared to normal lines. These new features impose higher demands on existing monitoring schemes. Therefore, this study proposed a method for monitoring ice shedding oscillations on transmission lines based on wireless attitude sensors. Firstly, the advantages of wireless pose sensors were analyzed,and a scheme was designed to install the sensor on the suspension insulator string of a straight tower, using the measured inclination angle combined with the string length to calculate the real-time span. Secondly, an algorithm was designed to solve the problem of difficult measurement of wire stiffness and real-time wire length, and combined with real-time span and real-time wire length to obtain the dynamic tension of the wire. Then, a transmission line monitoring algorithm based on the wire elastic deformation model is proposed, and a four towers three lines transmission line scaling experimental platform is built based on similarity criteria to verify the effectiveness of the model. The results show that this method effectively utilizes real-time data from attitude sensors to achieve precise real-time monitoring of tension and displacement on ice-shedding transmission lines, providing robust technical support for the safe and stable operation of power systems.关键词：pose sensor;frozen rain;transmission lines;analytical solution;real time span;elastic deformation30|26|0更新时间：2025-12-08
- 摘要：Due to the curse of dimensionality, effectively discarding redundant features while retaining critical information in high-dimensional data has become a key issue. Unsupervised feature selection, which performs dimensionality reduction without any prior class information, has attracted increasing attention. However, two common issues are ignored by existing unsupervised feature selection methods: Fuzziness is a common characteristic of data, but most existing unsupervised feature selection methods based on regularized regression ignore this aspect, resulting in suboptimal feature subsets; Most methods fail to effectively distinguish between normal and noisy samples and are susceptible to the noise. To tackle the mentioned issues, robust unsupervised feature selection with double fuzzy (DFRFS) learning is proposed. Specifically, DFRFS learning introduces fuzzy membership into unsupervised feature selection based on regularized regression, allowing data to be shared among multiple clusters, thereby better reflecting the complex structure and uncertainty of the data. Additionally, DFRFS learning assigns different weights to samples through the robust weight learning framework, thus suppressing the impact of noise while retaining the effect of normal samples. Experiments on toy and real-world datasets have demonstrated the effectiveness of the proposed method DFRFS learning.关键词：unsupervised learning;feature selection;fuzzy learning;sparse learning;dimensionality reduction47|29|0更新时间：2025-12-08
PAPERS
- 摘要：The physics of semiconductor devices, as a key scientific discipline for studying the working mechanisms and related physical phenomena of semiconductor devices, serves as a fundamental research direction in the field of semiconductor science and information devices. It integrates multidisciplinary theories from materials science, physics, chemistry, and microelectronics, providing theoretical support for advancing information technology and the semiconductor industry. This paper analyzes the application volume and funding success rates of various projects under the F0405 code in the field of semiconductor science and information devices over the past five years (2020—2024). Additionally, through a hotspot word cloud analysis of project titles and keywords from funded projects during this period, the paper summarizes the themes and trends in basic research on semiconductor device physics. The aim is to explore the funding characteristics of the national natural science foundation of China (NSFC) in the field of semiconductor device physics in recent years, providing insights for researchers in domestic academic institutions and enterprises to better understand the research hotspots, future development directions, and pathways in this field.关键词：semiconductor device physics;application and funding;research topics and hotspots;basic research;semiconductor science and information devices58|37|0更新时间：2025-12-10
- 摘要：One of the hottest topics in the field of wireless communication is ultra-reliable and low-latency communication (URLLC), where channel coding with short-and-medium length plays a critical role. Unlike the case for long codes, in the finite code length regime, the coding rate is constrained by the error performance, thus tailored coding construction, decoding algorithm design, and performance analysis and optimization are required. At present, there are studies on coding technologies such as polar codes and tail-biting convolutional codes for short-and-medium length codes, but they are mainly optimized designs for specific code lengths and rates, which are difficult to meet the requirements of flexible coding for practical applications. Based on this, this paper summarizes and discusses the state-of-the-art coding technology for short-and-medium length. Firstly, we review the theoretical bounds for the performance of finite-length codes. Then, we analyze the coding technologies proposed in recent years, and compare their advantages and disadvantages. Finally, we discuss in detail new coding technologies for URLLC, and we prospectively explore the future research directions and development trends of channel coding techniques for ultra-reliable and low-latency communication.关键词：channel coding;ultra-reliable and low-latency communication (URLLC);list decoding;finite-length capacity;polar code;tail-biting convolutional code (TBCC);twisted-pair superposition transmission (TPST)147|38|0更新时间：2025-12-08
SURVEYS AND REVIEWS
- 摘要：To decrease the impact of spontaneous emission noise and non-linear physical impairment on the request’s transmission performance in space division multiplexing elastic optical networks (SDM-EONs), an auxiliary graph-based bit loading and physical impairment-sensing resource allocation (BL-PIRA) method is proposed in the paper. In the BL-PIRA, the auxiliary graph is introduced to optimize the multi-policy routing selection with lowest cost for the request. Then, the bit loading mechanism is used to select fiber core and spectrum block that meet crosstalk between inter-cores and non-linear physical impairment for the request. The simulation results show that the proposed BL-PIRA can reduce the request’s blocking probability and improve spectrum utilization.关键词：SDM-EONs;physical impairment-sensing;auxiliary graph;bit-loading;spectrum utilization33|26|0更新时间：2025-12-08
CORRESPONDENCE
- Technical support is provided by Beijing Founder electronics co., LTD 京ICP备09064830号-19
京公网安备11010802024621 - It is recommended to read the content of this site in Chrome&IE9+. Please switch to extreme mode in browser 360.
- Cookies We use cookies to help provide and enhance our service and tailor content. By continuing, you agree to the use of cookies.
0