Lei Yang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Lei Yang has a total of 31 co-authors. Here are the top co-authors by number of shared publications:
- Guangyi Xiao
College of Computer Science and Electronic Engineering, Hunan University, Changsha, China
3 shared publications - Guoqi Xie
Hunan University
3 shared publications - Hao Chen
College of Computer Science and Electronic Engineering, Hunan University, Changsha, China
3 shared publications - Weiwei Xiang
School of Physics, Electronics and Intelligent Manufacturing, Huaihua University, China and Key Laboratory of Intelligent Control Technology for Wuling-Mountain Ecological Agriculture in Hunan Province, Huaihua University, China
3 shared publications - Xin Cheng
School of Information Science and Engineering, Hunan University, Changsha, Hunan, China
3 shared publications - Liming Ding
School of Computer and Artificial Intelligence (School of Software), Huaihua University, Huaihua 418000, China
2 shared publications
Latest Publications
- Open Access
- 9 October 2026
- Published by Elsevier in SSRN Electronic Journal
Dual-Stream Semantic Guidance with Prototype Anchor Calibration for Source-Fully-Free Adaptation of Vision-Language Models
Source-Fully-Free Domain Adaptation (SFF-DA) has emerged as a strategic paradigm to adapt Vision-Language Models (VLMs) without any access to source data or task-specific source models. However, we identify a critical Dual Semantic Drift that hinders this process: static drift arising from the rigidity of fixed class embeddings, and dynamic drift stemming from the divergence of generated captions, causing severe semantic misalignment that intensifies the stability-plasticity dilemma.To address this, we propose DSSG (Dual-Stream Semantic Guidance), an end-to-end framework that reconciles fine-grained plasticity with global stability. Our core contribution is the Dual Semantic Guidance (DSG) module, which integrates a caption stream for domain-specific knowledge with a class-anchor stream to anchor global categorical consistency. Furthermore, a Dynamic Cross-Modal Knowledge Distillation (CMKD) module is introduced to leverage the evolving teacher distribution for calibrating teacher-student consistency. Building upon DSSG, we further introduce Prototype Anchor Calibration (PAC), yielding DSSG-PAC, which periodically calibrates prototype anchors and caches them until the next calibration. This design reduces redundant text-side computation while preserving the adaptability of class guidance to the evolving text space. We further establish SFF-DA risk bounds that relate student risk to semantic-teacher quality and teacher--student discrepancy. Extensive experiments demonstrate that DSSG consistently outperforms current state-of-the-art methods across multiple benchmarks, while DSSG-PAC largely preserves its adaptation performance with 18.9% lower total adaptation time. The code is available at https://github.com/mrmenand/DSSG.
- 9 October 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessDual-Stream Semantic Guidance with Prototype Anchor Calibration for Source-Fully-Free Adaptation of Vision-Language Models
Source-Fully-Free Domain Adaptation (SFF-DA) has emerged as a strategic paradigm to adapt Vision-Language Models (VLMs) without any access to source data or task-specific source models. However, we identify a critical Dual Semantic Drift that hinders this process: static drift arising from the rigidity of fixed class embeddings, and dynamic drift stemming from the divergence of generated captions, causing severe semantic misalignment that intensifies the stability-plasticity dilemma.To address this, we propose DSSG (Dual-Stream Semantic Guidance), an end-to-end framework that reconciles fine-grained plasticity with global stability. Our core contribution is the Dual Semantic Guidance (DSG) module, which integrates a caption stream for domain-specific knowledge with a class-anchor stream to anchor global categorical consistency. Furthermore, a Dynamic Cross-Modal Knowledge Distillation (CMKD) module is introduced to leverage the evolving teacher distribution for calibrating teacher-student consistency. Building upon DSSG, we further introduce Prototype Anchor Calibration (PAC), yielding DSSG-PAC, which periodically calibrates prototype anchors and caches them until the next calibration. This design reduces redundant text-side computation while preserving the adaptability of class guidance to the evolving text space. We further establish SFF-DA risk bounds that relate student risk to semantic-teacher quality and teacher--student discrepancy. Extensive experiments demonstrate that DSSG consistently outperforms current state-of-the-art methods across multiple benchmarks, while DSSG-PAC largely preserves its adaptation performance with 18.9% lower total adaptation time. The code is available at https://github.com/mrmenand/DSSG.
- Open Access
- 13 July 2026
- Published by Elsevier in SSRN Electronic Journal
Cross-Modal Consensus Distillation with Prompt Adaptation for Source-Free Domain Adaptation
Source-Free Domain Adaptation (SFDA) aims to adapt a source-trained model to an unlabeled target domain without accessing source data. Recent unimodal-student + multimodal-teacher (U+M) methods introduce vision-language models (VLMs) as external teachers and transfer their semantic priors to the target model via cross-modal guidance and alignment. However, the reliability of the VLM branch in target domains remains insufficiently examined. We identify Bilateral Bias in U+M-based SFDA, where source-prior bias in the target model coexists with generic semantic bias in the VLM branch, yet their biased predictions are not fully aligned and can still provide exploitable cross-modal complementarity. Motivated by this observation, we propose Cross-Modal Consensus Distillation (CMCD), which builds a Product-of-Experts (PoE)-style cross-model consensus teacher via logit-level additive fusion and distills the consensus posterior into the target model. CMCD with Prompt Adaptation (CMCD-PA) further adapts the VLM branch through task-specific prompt adaptation, thereby improving the fused consensus posterior. A margin-level analysis reveals that the consensus gain of CMCD arises from heterogeneous complementarity rather than simple prediction addition. Extensive experiments across multiple benchmarks and CLIP teacher scales validate the effectiveness of CMCD and CMCD-PA while preserving lightweight inference. The code is available at https://github.com/mrmenand/CMCD.
- 13 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessCross-Modal Consensus Distillation with Prompt Adaptation for Source-Free Domain Adaptation
Source-Free Domain Adaptation (SFDA) aims to adapt a source-trained model to an unlabeled target domain without accessing source data. Recent unimodal-student + multimodal-teacher (U+M) methods introduce vision-language models (VLMs) as external teachers and transfer their semantic priors to the target model via cross-modal guidance and alignment. However, the reliability of the VLM branch in target domains remains insufficiently examined. We identify Bilateral Bias in U+M-based SFDA, where source-prior bias in the target model coexists with generic semantic bias in the VLM branch, yet their biased predictions are not fully aligned and can still provide exploitable cross-modal complementarity. Motivated by this observation, we propose Cross-Modal Consensus Distillation (CMCD), which builds a Product-of-Experts (PoE)-style cross-model consensus teacher via logit-level additive fusion and distills the consensus posterior into the target model. CMCD with Prompt Adaptation (CMCD-PA) further adapts the VLM branch through task-specific prompt adaptation, thereby improving the fused consensus posterior. A margin-level analysis reveals that the consensus gain of CMCD arises from heterogeneous complementarity rather than simple prediction addition. Extensive experiments across multiple benchmarks and CLIP teacher scales validate the effectiveness of CMCD and CMCD-PA while preserving lightweight inference. The code is available at https://github.com/mrmenand/CMCD.
- 18 January 2026
- Published by Elsevier in Expert Systems with Applications
- Vol. 309
ImCapDA: Fine-tuning CLIP via image captions for unsupervised domain adaptation
- 18 January 2026
- Published by Elsevier in Expert Systems with Applications
- Vol. 309
ImCapDA: Fine-tuning CLIP via image captions for unsupervised domain adaptation
- 10 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computational Social Systems
- Vol. 12 (5) , 3751-3764
Unsupervised Video Summarization Based on Spatiotemporal Semantic Graph and Enhanced Attention Mechanism
Generative adversarial networks (GANs) have demonstrated potential in enhancing keyframe selection and video reconstruction via adversarial training among unsupervised approaches. Nevertheless, GANs struggle to encapsulate the intricate spatiotemporal dynamics in videos, which is essential for producing coherent and informative summaries. To address these challenges, we introduce an unsupervised video summarization framework that synergistically integrates temporal–spatial semantic graphs (TSSGraphs) with a bilinear additive attention (BAA) mechanism. TSSGraphs are designed to effectively model temporal and spatial relationships among video frames by combining temporal convolution and dynamic edge convolution, thereby extracting salient features while mitigating model complexity. The BAA mechanism enhances the framework’s ability to capture critical motion information by addressing feature sparsity and eliminating redundant parameters, ensuring robust attention to significant motion dynamics. Experimental assessments on the SumMe and TVSum benchmark datasets reveal that our method attains improvements of up to 4.0% and 3.3% in F-score, respectively, compared to current methodologies. Moreover, our system demonstrates diminished parameter overhead throughout training and inference stages, particularly excelling in contexts with significant motion content.
- 10 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computational Social Systems
- Vol. 12 (5) , 3751-3764
Unsupervised Video Summarization Based on Spatiotemporal Semantic Graph and Enhanced Attention Mechanism
Generative adversarial networks (GANs) have demonstrated potential in enhancing keyframe selection and video reconstruction via adversarial training among unsupervised approaches. Nevertheless, GANs struggle to encapsulate the intricate spatiotemporal dynamics in videos, which is essential for producing coherent and informative summaries. To address these challenges, we introduce an unsupervised video summarization framework that synergistically integrates temporal–spatial semantic graphs (TSSGraphs) with a bilinear additive attention (BAA) mechanism. TSSGraphs are designed to effectively model temporal and spatial relationships among video frames by combining temporal convolution and dynamic edge convolution, thereby extracting salient features while mitigating model complexity. The BAA mechanism enhances the framework’s ability to capture critical motion information by addressing feature sparsity and eliminating redundant parameters, ensuring robust attention to significant motion dynamics. Experimental assessments on the SumMe and TVSum benchmark datasets reveal that our method attains improvements of up to 4.0% and 3.3% in F-score, respectively, compared to current methodologies. Moreover, our system demonstrates diminished parameter overhead throughout training and inference stages, particularly excelling in contexts with significant motion content.
- 13 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Knowledge and Data Engineering
- Vol. 37 (9) , 5330-5345
Intra-Trajectory Error Balancing and Inter-Trajectory Feature Point Clustering for Trajectory Compression
The widespread use of locatable devices leads to a sharp increase in the storage of trajectory data, and redundant storage of similar trajectories wastes a large amount of storage resources. The state-of-the-art multiple trajectory compression algorithms are developed to strip the partial information of trajectory; however, these algorithms have low compression efficiency because they do not eliminate the redundancy within a single trajectory as much as possible, as well as high time overhead due to matching of reference sub-trajectories. In this study, we propose a new spatio-temporal trajectory compression technique, consisting of intra-trajectory error balancing and inter-trajectory feature point clustering. Intra-trajectory error balancing is achieved through retaining high score (an aggregated metric) trajectory points (i.e., feature points). Furthermore, inter-trajectory feature point clustering realizes the fusion of similar trajectories and extracts the commonality between trajectories. Experiments are performed on five real trajectory datasets, including two road datasets, one airline dataset, and one walking dataset. Compared with the state-of-the-art methods, our compression technique improves the compression ratio by an average of 24.9% under the same error, and reduces the time overhead by at least an order of magnitude.
- 13 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Knowledge and Data Engineering
- Vol. 37 (9) , 5330-5345
Intra-Trajectory Error Balancing and Inter-Trajectory Feature Point Clustering for Trajectory Compression
The widespread use of locatable devices leads to a sharp increase in the storage of trajectory data, and redundant storage of similar trajectories wastes a large amount of storage resources. The state-of-the-art multiple trajectory compression algorithms are developed to strip the partial information of trajectory; however, these algorithms have low compression efficiency because they do not eliminate the redundancy within a single trajectory as much as possible, as well as high time overhead due to matching of reference sub-trajectories. In this study, we propose a new spatio-temporal trajectory compression technique, consisting of intra-trajectory error balancing and inter-trajectory feature point clustering. Intra-trajectory error balancing is achieved through retaining high score (an aggregated metric) trajectory points (i.e., feature points). Furthermore, inter-trajectory feature point clustering realizes the fusion of similar trajectories and extracts the commonality between trajectories. Experiments are performed on five real trajectory datasets, including two road datasets, one airline dataset, and one walking dataset. Compared with the state-of-the-art methods, our compression technique improves the compression ratio by an average of 24.9% under the same error, and reduces the time overhead by at least an order of magnitude.
- 31 July 2024
- Published by Springer Nature
- p. 405-419
Fetal Cardiac Structure Detection Using Multi-task Learning
- 31 July 2024
- Published by Springer Nature
- p. 405-419
Fetal Cardiac Structure Detection Using Multi-task Learning
- 21 February 2024
- Published by Elsevier in Neurocomputing
- Vol. 579
Fetal cardiac ultrasound standard section detection model based on multitask learning and mixed attention mechanism
- 21 February 2024
- Published by Elsevier in Neurocomputing
- Vol. 579
Fetal cardiac ultrasound standard section detection model based on multitask learning and mixed attention mechanism
- 7 November 2023
- Published by Elsevier in Expert Systems with Applications
- Vol. 240, 122384
HAMFace: Hardness adaptive margin loss for face recognition with various intra-class variations
- 7 November 2023
- Published by Elsevier in Expert Systems with Applications
- Vol. 240, 122384
HAMFace: Hardness adaptive margin loss for face recognition with various intra-class variations
- 26 October 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Parallel and Distributed Systems
- Vol. 34 (12) , 3309-3322
Efficient Edge Data Management Framework for IIoT via Prediction-Based Data Reduction
Large amounts of time series data are required to support data analysis at the edge in the end-edge-cloud Industrial Internet of Things (IIoT) architecture. Reducing the storage cost is one of the main challenges in edge data management due to the limited storage resource of edge nodes. The state-of-the-art data reduction method has a high time overhead and poor reduction efficiency for unstable data sets. To solve this problem, this study proposes a time-series data management framework that combines data partition and data compression techniques. For the data partition technique, we propose an adaptive selection strategy to integrate the access pattern of the application and the characteristics of the time series data, thereby improving the partition accuracy. For the data compression technique, we propose a compression scheme based on time series data segmentation by using the idea of divide and conquer; we further introduce a change point detection technique to improve the compression efficiency for unstable data sets. Experimental results obtained with three types of real industrial data sets show that our framework is significantly better than the state-of-the-art method in terms of compression ratio and time overhead.
- 26 October 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Parallel and Distributed Systems
- Vol. 34 (12) , 3309-3322
Efficient Edge Data Management Framework for IIoT via Prediction-Based Data Reduction
Large amounts of time series data are required to support data analysis at the edge in the end-edge-cloud Industrial Internet of Things (IIoT) architecture. Reducing the storage cost is one of the main challenges in edge data management due to the limited storage resource of edge nodes. The state-of-the-art data reduction method has a high time overhead and poor reduction efficiency for unstable data sets. To solve this problem, this study proposes a time-series data management framework that combines data partition and data compression techniques. For the data partition technique, we propose an adaptive selection strategy to integrate the access pattern of the application and the characteristics of the time series data, thereby improving the partition accuracy. For the data compression technique, we propose a compression scheme based on time series data segmentation by using the idea of divide and conquer; we further introduce a change point detection technique to improve the compression efficiency for unstable data sets. Experimental results obtained with three types of real industrial data sets show that our framework is significantly better than the state-of-the-art method in terms of compression ratio and time overhead.
- 3 October 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Parallel and Distributed Systems
- Vol. 34 (12) , 3090-3103
OfpCNN: On-Demand Fine-Grained Partitioning for CNN Inference Acceleration in Heterogeneous Devices
Collaborative inference is a promising method for balancing the limited computational power of Internet of Things (IoT) devices with the huge computational demands of convolutional neural networks (CNNs). In this approach, a CNN is divided into multiple partitions and placed on multiple devices to run simultaneously. However, two major challenges are raised. (1) Computational latencies vary when the central processing unit (CPU) loads of devices are different. However, no suitable methods are available for accurately determining computation latencies on the basis of CPU utilization. (2) Existing methods partition a CNN model either vertically or horizontally. The granularity of these methods is extremely coarse and their accuracy is low. To address the aforementioned issues, this study proposes a distributed collaborative inference framework that supports a fine-grained partitioning scheme for CNN in heterogeneous devices (hereafter referred to as OfpCNN). First, the framework uses the layer latency prediction model based on floating-point operations and CPU load (FCPM) to accurately predict the computation latency of each layer of CNN in different devices. Subsequently, OfpCNN uses horizontal and vertical partitioning methods (HVPM) to partition the input feature maps and the structure of CNN respectively in accordance with network conditions and computing capacity, then assigns them to multiple devices for execution. The HVPM solution overall considers the execution position of the layer, parallelism, and location of devices responsible for data aggregation and distribution, which can consequently obtain more fine-grained partition schemes. Experimental results show that FCPM can achieve a minimum accuracy of 88% and HVPM can improve the inference speed by 1-2.54 times compared with other state-of-the-art methods.
- 3 October 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Parallel and Distributed Systems
- Vol. 34 (12) , 3090-3103
OfpCNN: On-Demand Fine-Grained Partitioning for CNN Inference Acceleration in Heterogeneous Devices
Collaborative inference is a promising method for balancing the limited computational power of Internet of Things (IoT) devices with the huge computational demands of convolutional neural networks (CNNs). In this approach, a CNN is divided into multiple partitions and placed on multiple devices to run simultaneously. However, two major challenges are raised. (1) Computational latencies vary when the central processing unit (CPU) loads of devices are different. However, no suitable methods are available for accurately determining computation latencies on the basis of CPU utilization. (2) Existing methods partition a CNN model either vertically or horizontally. The granularity of these methods is extremely coarse and their accuracy is low. To address the aforementioned issues, this study proposes a distributed collaborative inference framework that supports a fine-grained partitioning scheme for CNN in heterogeneous devices (hereafter referred to as OfpCNN). First, the framework uses the layer latency prediction model based on floating-point operations and CPU load (FCPM) to accurately predict the computation latency of each layer of CNN in different devices. Subsequently, OfpCNN uses horizontal and vertical partitioning methods (HVPM) to partition the input feature maps and the structure of CNN respectively in accordance with network conditions and computing capacity, then assigns them to multiple devices for execution. The HVPM solution overall considers the execution position of the layer, parallelism, and location of devices responsible for data aggregation and distribution, which can consequently obtain more fine-grained partition schemes. Experimental results show that FCPM can achieve a minimum accuracy of 88% and HVPM can improve the inference speed by 1-2.54 times compared with other state-of-the-art methods.