Zhao Lv
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Zhao Lv has a total of 227 co-authors. Here are the top co-authors by number of shared publications:
- Cunhang Fan
State Key Laboratory of Opto-Electronic Information Acquisition and Protection Technology, School of Computer Science and Technology, Anhui University, Hefei, Anhui, China
33 shared publications - Jianhua Tao
Tsinghua University, Beijing, China
13 shared publications - Minchao Wu
Hefei Normal University
10 shared publications - Xinhui Li
State Key Laboratory of Opto-Electronic Information Acquisition and Protection Technology, (School of Computer Science and Technology), Anhui University, Hefei, China
10 shared publications - Jian Zhou
Anhui University
9 shared publications - Ping Li
School of Computer Science and Technology, Anhui University
7 shared publications
Latest Publications
- Open Access
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 17749-17756
Beyond Speech: Dual-Domain SSL Fusion for Unified All-Type Audio Deepfake Detection
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 17749-17756
Open AccessBeyond Speech: Dual-Domain SSL Fusion for Unified All-Type Audio Deepfake Detection
- 6 November 2026
- Published by Springer Nature
- p. 459-473
VCLSTM: A Novel Deep Learning Framework for EOG-Based Emotion Recognition via Variational Mode Decomposition
- 6 November 2026
- Published by Springer Nature
- p. 459-473
VCLSTM: A Novel Deep Learning Framework for EOG-Based Emotion Recognition via Variational Mode Decomposition
- Open Access
- 9 September 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Autonomous and Adaptive Systems
DA2Mel: A Dual-Dynamic Adaptive Network for EEG-to-Mel Spectrogram Reconstruction
Decoding neural activity into natural speech is a frontier in cognitive neuroscience and brain-computer interfaces. Existing methods overlook two critical issues when decoding continuous long-sequence speech: the pronounced inter-subject variability and substantial temporal fluctuations within individual EEG signals, as well as the inherent transverse sparsity present in acoustic features. To address these issues, this paper proposes DA2Mel, a dual-dynamic adaptive network that enhances the perception of non-stationary neural signals. The Dynamic Subject-Specific Memory integrator (DSSM) forms an adaptive closed loop using identity-aware memory and momentum updating. It adapts to EEG distribution shifts from inter- and intra-subject variations, aligns with dynamic physiological states, and ensures long-term robustness. The Dual-Pathway Spatial-Channel Processor (DP-SCP) extracts critical channel information, fuses multi-scale spatial features, and bridges global trends and local details. In addition, the Stacked Fine-Grained Module (SFGM) employs a content-aware adaptive mechanism, fusing local convolutions with partial Top- \(k\) sparse attention. By dynamically reconstructing representation weights based on the instantaneous sparsity of acoustic features, SFGM distills pivotal characteristics and optimizes semantics under noise. Experimental results on the SparrKULee dataset demonstrate that DA2Mel outperforms state-of-the-art methods in both the Held-out Stories and Held-out Subjects settings, validating the superiority of our framework for continuous speech decoding.- 9 September 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Autonomous and Adaptive Systems
Open AccessDA2Mel: A Dual-Dynamic Adaptive Network for EEG-to-Mel Spectrogram Reconstruction
Decoding neural activity into natural speech is a frontier in cognitive neuroscience and brain-computer interfaces. Existing methods overlook two critical issues when decoding continuous long-sequence speech: the pronounced inter-subject variability and substantial temporal fluctuations within individual EEG signals, as well as the inherent transverse sparsity present in acoustic features. To address these issues, this paper proposes DA2Mel, a dual-dynamic adaptive network that enhances the perception of non-stationary neural signals. The Dynamic Subject-Specific Memory integrator (DSSM) forms an adaptive closed loop using identity-aware memory and momentum updating. It adapts to EEG distribution shifts from inter- and intra-subject variations, aligns with dynamic physiological states, and ensures long-term robustness. The Dual-Pathway Spatial-Channel Processor (DP-SCP) extracts critical channel information, fuses multi-scale spatial features, and bridges global trends and local details. In addition, the Stacked Fine-Grained Module (SFGM) employs a content-aware adaptive mechanism, fusing local convolutions with partial Top- \(k\) sparse attention. By dynamically reconstructing representation weights based on the instantaneous sparsity of acoustic features, SFGM distills pivotal characteristics and optimizes semantics under noise. Experimental results on the SparrKULee dataset demonstrate that DA2Mel outperforms state-of-the-art methods in both the Held-out Stories and Held-out Subjects settings, validating the superiority of our framework for continuous speech decoding. - Open Access
- 19 August 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Autonomous and Adaptive Systems
MS²FL: Modality-shared and Modality-specific Feature Learning for Multimodal Emotion Recognition
Multimodal emotion recognition based on complementary physiological signals such as electroencephalogram (EEG) and eye movements can effectively reflect human emotional states, demonstrating significant potential in fields such as rehabilitation monitoring and driving safety. However, existing multimodal emotion recognition approaches often focus solely on either feature fusion or feature alignment strategies, lacking a unified framework capturing both complementary and modality-specific representations. To address this limitation, we propose a modality-shared and modality-specific feature learning framework (MS \({}^{2}\) FL) for multimodal emotion recognition. Specifically, a dual-path encoder is employed first to extract the spatiotemporal frequency features of EEG signals and dynamic response features of eye movement signals, respectively. Then, a specificity-preserving feature distribution alignment mechanism is introduced to alleviate distribution discrepancies between modalities while reinforcing modality-specific modeling of distinctive features. Finally, a feature distribution enhancement fusion strategy is utilized to effectively integrate the shared features across modalities, thus improving the representational capacity of the fused features. Experimental results demonstrate that MS \({}^{2}\) FL achieves recognition accuracies of 96.99% and 89.83% on the SEED and SEED-V datasets, respectively, significantly improving emotion recognition performance. This study provides a concise and effective solution for emotion perception neurotechnology and offers a practical solution for emotion perception health monitoring and rehabilitation.- 19 August 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Autonomous and Adaptive Systems
Open AccessMS²FL: Modality-shared and Modality-specific Feature Learning for Multimodal Emotion Recognition
Multimodal emotion recognition based on complementary physiological signals such as electroencephalogram (EEG) and eye movements can effectively reflect human emotional states, demonstrating significant potential in fields such as rehabilitation monitoring and driving safety. However, existing multimodal emotion recognition approaches often focus solely on either feature fusion or feature alignment strategies, lacking a unified framework capturing both complementary and modality-specific representations. To address this limitation, we propose a modality-shared and modality-specific feature learning framework (MS \({}^{2}\) FL) for multimodal emotion recognition. Specifically, a dual-path encoder is employed first to extract the spatiotemporal frequency features of EEG signals and dynamic response features of eye movement signals, respectively. Then, a specificity-preserving feature distribution alignment mechanism is introduced to alleviate distribution discrepancies between modalities while reinforcing modality-specific modeling of distinctive features. Finally, a feature distribution enhancement fusion strategy is utilized to effectively integrate the shared features across modalities, thus improving the representational capacity of the fused features. Experimental results demonstrate that MS \({}^{2}\) FL achieves recognition accuracies of 96.99% and 89.83% on the SEED and SEED-V datasets, respectively, significantly improving emotion recognition performance. This study provides a concise and effective solution for emotion perception neurotechnology and offers a practical solution for emotion perception health monitoring and rehabilitation. - 1 July 2026
- Published by IOP Publishing in Biomedical Physics & Engineering Express
- Vol. 12  (4) , 045033
MELF: A multi-view ensemble learning framework for normative resting state EEG signal quality assessment
Objective. Normative resting-state electroencephalography (rsEEG), defined as standardized recordings from healthy populations, provides essential baselines for studying brain mechanisms and psychiatric disorders. However, its reliability critically depends on rigorous quality assessment, as artifacts or over-processing can distort subsequent analyses. Current approaches often rely on single-feature perspectives, which fail to capture the multidimensional nature of signal quality. Approach. We proposes MELF, a quality assessment framework for normative rsEEG that systematically integrates three complementary feature domains: time domain (mobility, complexity), frequency domain (α-band relative power, parallel log spectra index), and spatial domain (α-band directed phase lag index, bad channel ratio, ICLabel-classified brain component proportion). A random forest-based ensemble classifier was constructed using dynamic weight allocation based on cross-validation accuracy. Main results. Experiments on four open rsEEG datasets (DSTR, TRCS, LEMON, HBN) and expert-annotated clinical data demonstrated that: (1) MELF achieved up to 10% accuracy improvement over single-view methods; (2) MELF outperformed traditional machine learning and deep learning baselines while reducing computational time, reaching 97.69% accuracy (AUC = 0.9957), and maintained strong performance on an independent test set (92% accuracy, AUC = 0.88); (3) the spectral view contributed the most, confirming the evaluative capability of the parallel log spectra index (PaLOSi). Significance. MELF provides an interpretable, high-accuracy, and robust automated framework for normative rsEEG quality assessment, with broad applications in neuroscience research and clinical quality control.
- 1 July 2026
- Published by IOP Publishing in Biomedical Physics & Engineering Express
- Vol. 12  (4) , 045033
MELF: A multi-view ensemble learning framework for normative resting state EEG signal quality assessment
Objective. Normative resting-state electroencephalography (rsEEG), defined as standardized recordings from healthy populations, provides essential baselines for studying brain mechanisms and psychiatric disorders. However, its reliability critically depends on rigorous quality assessment, as artifacts or over-processing can distort subsequent analyses. Current approaches often rely on single-feature perspectives, which fail to capture the multidimensional nature of signal quality. Approach. We proposes MELF, a quality assessment framework for normative rsEEG that systematically integrates three complementary feature domains: time domain (mobility, complexity), frequency domain (α-band relative power, parallel log spectra index), and spatial domain (α-band directed phase lag index, bad channel ratio, ICLabel-classified brain component proportion). A random forest-based ensemble classifier was constructed using dynamic weight allocation based on cross-validation accuracy. Main results. Experiments on four open rsEEG datasets (DSTR, TRCS, LEMON, HBN) and expert-annotated clinical data demonstrated that: (1) MELF achieved up to 10% accuracy improvement over single-view methods; (2) MELF outperformed traditional machine learning and deep learning baselines while reducing computational time, reaching 97.69% accuracy (AUC = 0.9957), and maintained strong performance on an independent test set (92% accuracy, AUC = 0.88); (3) the spectral view contributed the most, confirming the evaluative capability of the parallel log spectra index (PaLOSi). Significance. MELF provides an interpretable, high-accuracy, and robust automated framework for normative rsEEG quality assessment, with broad applications in neuroscience research and clinical quality control.
- Open Access
- 13 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-19
A Shared Look: Detecting Deepfakes with Inter-Subject Neural Synchrony
- 13 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-19
Open AccessA Shared Look: Detecting Deepfakes with Inter-Subject Neural Synchrony
- 9 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. PP  (00983063) , 1
CMA-DF: A physiological signal fusion method based on cross-modal alignment and dynamic fusion
Multimodal emotion recognition based on complementary physiological signals, such as an electroencephalogram (EEG) and eye movement, can effectively reflect human emotional states, showing great potential in fields like rehabilitation monitoring and driving safety management. However, existing methods usually directly map multimodal features into a unified feature space for information fusion, which not only fails to eliminate the heterogeneity between different modalities but also ignores the impact of the modal contribution on the fusion results. To address these limitations, this paper proposes a cross-modal alignment and dynamic fusion (CMA-DF) learning framework to improve the performance of multimodal emotion recognition. Specifically, we first adopt a two-branch feature extractor to respectively extract robust feature representations of EEG signals and eye movement signals. Secondly, a cross-modal alignment network is designed, which reduces distribution differences through collaborative modeling of intermodal interaction relationships and alignment strategies, thus providing a consistent basis for subsequent fusion. Subsequently, a dynamic fusion strategy is introduced to efficiently integrate complementary information from multiple modalities while adaptively balancing the contribution weights of each modality, thereby enhancing the representational capability of the fused features. Finally, experimental results in the SEED and SEED-V datasets show that the proposed method achieves recognition accuracies of 96.69% and 88.03%, respectively. Comparative experiments fully verify its superiority in multimodal feature alignment, fusion, and emotion recognition performance.
- 9 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. PP  (00983063) , 1
CMA-DF: A physiological signal fusion method based on cross-modal alignment and dynamic fusion
Multimodal emotion recognition based on complementary physiological signals, such as an electroencephalogram (EEG) and eye movement, can effectively reflect human emotional states, showing great potential in fields like rehabilitation monitoring and driving safety management. However, existing methods usually directly map multimodal features into a unified feature space for information fusion, which not only fails to eliminate the heterogeneity between different modalities but also ignores the impact of the modal contribution on the fusion results. To address these limitations, this paper proposes a cross-modal alignment and dynamic fusion (CMA-DF) learning framework to improve the performance of multimodal emotion recognition. Specifically, we first adopt a two-branch feature extractor to respectively extract robust feature representations of EEG signals and eye movement signals. Secondly, a cross-modal alignment network is designed, which reduces distribution differences through collaborative modeling of intermodal interaction relationships and alignment strategies, thus providing a consistent basis for subsequent fusion. Subsequently, a dynamic fusion strategy is introduced to efficiently integrate complementary information from multiple modalities while adaptively balancing the contribution weights of each modality, thereby enhancing the representational capability of the fused features. Finally, experimental results in the SEED and SEED-V datasets show that the proposed method achieves recognition accuracies of 96.69% and 88.03%, respectively. Comparative experiments fully verify its superiority in multimodal feature alignment, fusion, and emotion recognition performance.
- Open Access
TMacaque-FaceNet: Automatic Facial Recognition Based on Vision Transformer for Wild Tibetan Macaques
Within the framework of behavioral ecology and conservation, individual recognition plays a critical role in the research on wild social animals at the individual level. Traditional identification methods often rely on long-term field experience or invasive physical tagging. Recent advances in deep learning enable non-invasive individual recognition under natural conditions; however, the effectiveness of facial detection and identification depends on species-specific facial characteristics, environmental conditions, and dataset scale. In this study, we used 3385 images from 18 identified wild Tibetan macaques (Macaca thibetana) to develop an individual recognition system, TMacaque-FaceNet, integrating You Only Look Once (YOLO) for face detection and a Vision Transformer (ViT) for individual classification. The results showed that the Tibetan macaque face detector achieved a mAP@0.5 of 0.971, with a precision of 0.974 and a recall of 0.931. The individual recognizer for the wild Tibetan macaque social group achieved a top-1 accuracy of 96.33% on the test set. On an event-wise (temporal holdout) validation set comprising 90 images (5 images per individual), the recognizer achieved a top-1 accuracy of 95.56%. Gradient-weighted attention rollout analyses further revealed that the model focused on biologically meaningful facial regions, supporting the interpretability of the recognition process. Our results provide a new automated facial recognition method to non-invasively monitor Tibetan macaque individuals in natural environments. It provides a practical tool to facilitate automated behavioral observation, social network analysis, and long-term population monitoring of wild non-human primates.
Open AccessTMacaque-FaceNet: Automatic Facial Recognition Based on Vision Transformer for Wild Tibetan Macaques
Within the framework of behavioral ecology and conservation, individual recognition plays a critical role in the research on wild social animals at the individual level. Traditional identification methods often rely on long-term field experience or invasive physical tagging. Recent advances in deep learning enable non-invasive individual recognition under natural conditions; however, the effectiveness of facial detection and identification depends on species-specific facial characteristics, environmental conditions, and dataset scale. In this study, we used 3385 images from 18 identified wild Tibetan macaques (Macaca thibetana) to develop an individual recognition system, TMacaque-FaceNet, integrating You Only Look Once (YOLO) for face detection and a Vision Transformer (ViT) for individual classification. The results showed that the Tibetan macaque face detector achieved a mAP@0.5 of 0.971, with a precision of 0.974 and a recall of 0.931. The individual recognizer for the wild Tibetan macaque social group achieved a top-1 accuracy of 96.33% on the test set. On an event-wise (temporal holdout) validation set comprising 90 images (5 images per individual), the recognizer achieved a top-1 accuracy of 95.56%. Gradient-weighted attention rollout analyses further revealed that the model focused on biologically meaningful facial regions, supporting the interpretability of the recognition process. Our results provide a new automated facial recognition method to non-invasively monitor Tibetan macaque individuals in natural environments. It provides a practical tool to facilitate automated behavioral observation, social network analysis, and long-term population monitoring of wild non-human primates.
- Top Cited Papers
- 25 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Audio, Speech and Language Processing
- Vol. 34, 2077-2089
SSE-Net: Toward Low-Power-Consumption Spiking Neural Network for Monaural Speech Enhancement
Speech enhancement (SE) as a front-end task, especially in edge device deployment, requires low complexity and power consumption. However, deploying most artificial neural network (ANN)-based SE models while ensuring performance is challenging, particularly with the rapid development of large models in recent years. Spiking neural networks (SNNs) have shown potential in reducing power consumption. However, SNN-based SE models face two main challenges: on one hand, directly converting an ANN model to a SNN model introduces redundant modules, leading to difficulties in model training and information mismatch; on the other hand, the discrete binary activation and complex spatio-temporal dynamics of SNNs often result in information loss. To address these challenges, this paper proposes the SNN Speech Enhancement Network (SSE-Net). Unlike conventional approaches that convert ANNs to SNNs, all modules in SSE-Net are specifically designed for spike signal characteristics. Furthermore, the method innovatively develops a spiking feature extraction group that simultaneously converts information into spike signals while extracting original speech features, thereby effectively reducing information loss while significantly lowering both power consumption and computational complexity. The proposed method further designs information transformation blocks to convert spike signals back into continuous signals, fine-tuning and supplementing the speech information. Experimental results demonstrate that our proposed model achieves state of the art (SOTA) performance compared to the best SNN-SE models, with a 62% reduction in power proxy and 11% reduction in power cost, it also demonstrates strong competitiveness when compared with ANN-based SE models.
- 25 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Audio, Speech and Language Processing
- Vol. 34, 2077-2089
Top Cited PapersSSE-Net: Toward Low-Power-Consumption Spiking Neural Network for Monaural Speech Enhancement
Speech enhancement (SE) as a front-end task, especially in edge device deployment, requires low complexity and power consumption. However, deploying most artificial neural network (ANN)-based SE models while ensuring performance is challenging, particularly with the rapid development of large models in recent years. Spiking neural networks (SNNs) have shown potential in reducing power consumption. However, SNN-based SE models face two main challenges: on one hand, directly converting an ANN model to a SNN model introduces redundant modules, leading to difficulties in model training and information mismatch; on the other hand, the discrete binary activation and complex spatio-temporal dynamics of SNNs often result in information loss. To address these challenges, this paper proposes the SNN Speech Enhancement Network (SSE-Net). Unlike conventional approaches that convert ANNs to SNNs, all modules in SSE-Net are specifically designed for spike signal characteristics. Furthermore, the method innovatively develops a spiking feature extraction group that simultaneously converts information into spike signals while extracting original speech features, thereby effectively reducing information loss while significantly lowering both power consumption and computational complexity. The proposed method further designs information transformation blocks to convert spike signals back into continuous signals, fine-tuning and supplementing the speech information. Experimental results demonstrate that our proposed model achieves state of the art (SOTA) performance compared to the best SNN-SE models, with a 62% reduction in power proxy and 11% reduction in power cost, it also demonstrates strong competitiveness when compared with ANN-based SE models.
- Open Access
FAIR-ESI: Feature Adaptive Importance Refinement for Electrophysiological Source Imaging
An essential technique for diagnosing brain disorders is electrophysiological source imaging (ESI). While model-based optimization and deep learning methods have achieved promising results in this field, the accurate selection and refinement of features remains a central challenge for precise ESI. This paper proposes FAIR-ESI, a novel framework that adaptively refines feature importance across different views, including FFT-based spectral feature refinement, weighted temporal feature refinement, and self-attention-based patch-wise feature refinement. Extensive experiments on two simulation datasets with diverse configurations and two real-world clinical datasets validate our framework's efficacy, highlighting its potential to advance brain disorder diagnosis and offer new insights into brain function.
Open AccessFAIR-ESI: Feature Adaptive Importance Refinement for Electrophysiological Source Imaging
An essential technique for diagnosing brain disorders is electrophysiological source imaging (ESI). While model-based optimization and deep learning methods have achieved promising results in this field, the accurate selection and refinement of features remains a central challenge for precise ESI. This paper proposes FAIR-ESI, a novel framework that adaptively refines feature importance across different views, including FFT-based spectral feature refinement, weighted temporal feature refinement, and self-attention-based patch-wise feature refinement. Extensive experiments on two simulation datasets with diverse configurations and two real-world clinical datasets validate our framework's efficacy, highlighting its potential to advance brain disorder diagnosis and offer new insights into brain function.