Zhaohui Cai
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Zhaohui Cai has a total of 37 co-authors. Here are the top co-authors by number of shared publications:
- Shubo Liu
Wuhan University
8 shared publications - Guoqing Tu
School of Cyber Science and Engineering, Wuhan University, Wuhan, China
5 shared publications Guoqing Tu
Wuhan University
3 shared publicationsShubo Liu
Wuhan University
3 shared publications- Zhuoqing Chang
School of Computer Science, Wuhan University, Wuhan, China
3 shared publications - Chang Xiong
School of Computer Science, Wuhan University, Wuhan, China
2 shared publications
Latest Publications
- 18 November 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 12 (7) , 8552-8569
Continuous Latent Adversarial Autoencoder: A Time-Sensitive Method for Incomplete Time-Series Modeling
Incomplete time-series modeling is an unavoidable topic in real-world time-series analysis because of the frequent occurrence of missing values in practical data. However, integrating data preprocessing and subsequent analysis within a model can amplify the errors from processed values. Moreover, most existing methods that directly model incomplete time series often fail to infer values at any desired time or support multistep prediction. To address these issues, this article introduces a novel generative model called the continuous latent adversarial autoencoder (CLAAE) for directly modeling incomplete time series. CLAAE can effectively impute missing data of any time point and support multistep prediction. Specifically, CLAAE devises a time-aware long short-term memory (LSTM) encoder to extract temporal and sequential characteristics. The decoder is built upon the augmented neural ordinary differential equation (ANODE), allowing it to infer the probability of missing data across an arbitrary continuous-time horizon. To guarantee the meaningfulness of samples generated from any region within the prior space, a fully connected neural network is utilized as a discriminator, encouraging the aggregated posterior learned by the encoder to be indistinguishable from a selected prior distribution. Extensive experimental results across simulations and real-world datasets demonstrate that CLAAE outperforms baseline methods, especially when the amount of missing data is overwhelming. By combining the autoencoder and adversarial training, CLAAE can significantly enhance the quality of the synthetic samples, respecting the original feature distributions and the temporal dynamics.
- 18 November 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 12 (7) , 8552-8569
Continuous Latent Adversarial Autoencoder: A Time-Sensitive Method for Incomplete Time-Series Modeling
Incomplete time-series modeling is an unavoidable topic in real-world time-series analysis because of the frequent occurrence of missing values in practical data. However, integrating data preprocessing and subsequent analysis within a model can amplify the errors from processed values. Moreover, most existing methods that directly model incomplete time series often fail to infer values at any desired time or support multistep prediction. To address these issues, this article introduces a novel generative model called the continuous latent adversarial autoencoder (CLAAE) for directly modeling incomplete time series. CLAAE can effectively impute missing data of any time point and support multistep prediction. Specifically, CLAAE devises a time-aware long short-term memory (LSTM) encoder to extract temporal and sequential characteristics. The decoder is built upon the augmented neural ordinary differential equation (ANODE), allowing it to infer the probability of missing data across an arbitrary continuous-time horizon. To guarantee the meaningfulness of samples generated from any region within the prior space, a fully connected neural network is utilized as a discriminator, encouraging the aggregated posterior learned by the encoder to be indistinguishable from a selected prior distribution. Extensive experimental results across simulations and real-world datasets demonstrate that CLAAE outperforms baseline methods, especially when the amount of missing data is overwhelming. By combining the autoencoder and adversarial training, CLAAE can significantly enhance the quality of the synthetic samples, respecting the original feature distributions and the temporal dynamics.
- 1 October 2024
- Published by Springer Nature
- p. 409-424
VFDV-IM: An Efficient and Securely Vertical Federated Data Valuation
- 1 October 2024
- Published by Springer Nature
- p. 409-424
VFDV-IM: An Efficient and Securely Vertical Federated Data Valuation
- 17 September 2024
- Published by Springer Nature
- p. 175-189
Missing Data Imputation via Neighbor Data Feature-Enriched Neural Ordinary Differential Equations
- 17 September 2024
- Published by Springer Nature
- p. 175-189
Missing Data Imputation via Neighbor Data Feature-Enriched Neural Ordinary Differential Equations
- 12 March 2024
- Published by Springer Nature
- p. 439-454
A Time Series Data Compression Co-processor Based on RISC-V Custom Instructions
- 12 March 2024
- Published by Springer Nature
- p. 439-454
A Time Series Data Compression Co-processor Based on RISC-V Custom Instructions
- 2 August 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Image Processing
- Vol. 32 (10577149) , 4543-4554
Composed Image Retrieval via Cross Relation Network With Hierarchical Aggregation Transformer
Composing Text and Image to Image Retrieval (CTI-IR) aims at finding the target image, which matches the query image visually along with the query text semantically. However, existing works ignore the fact that the reference text usually serves multiple functions, e.g., modification and auxiliary. To address this issue, we put forth a unified solution, namely Hierarchical Aggregation Transformer incorporated with Cross Relation Network (CRN). CRN unifies modification and relevance manner in a single framework. This configuration shows broader applicability, enabling us to model both modification and auxiliary text or their combination in triplet relationships simultaneously. Specifically, CRN includes: 1) Cross Relation Network comprehensively captures the relationships of various composed retrieval scenarios caused by two different query text types, allowing a unified retrieval model to designate adaptive combination strategies for flexible applicability; 2) Hierarchical Aggregation Transformer aggregates top-down features with Multi-layer Perceptron (MLP) to overcome the limitations of edge information loss in a window-based multi-stage Transformer. Extensive experiments demonstrate the superiority of the proposed CRN over all three fashion-domain datasets. Code is available at github.com/yan9qu/crn.
- 2 August 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Image Processing
- Vol. 32 (10577149) , 4543-4554
Composed Image Retrieval via Cross Relation Network With Hierarchical Aggregation Transformer
Composing Text and Image to Image Retrieval (CTI-IR) aims at finding the target image, which matches the query image visually along with the query text semantically. However, existing works ignore the fact that the reference text usually serves multiple functions, e.g., modification and auxiliary. To address this issue, we put forth a unified solution, namely Hierarchical Aggregation Transformer incorporated with Cross Relation Network (CRN). CRN unifies modification and relevance manner in a single framework. This configuration shows broader applicability, enabling us to model both modification and auxiliary text or their combination in triplet relationships simultaneously. Specifically, CRN includes: 1) Cross Relation Network comprehensively captures the relationships of various composed retrieval scenarios caused by two different query text types, allowing a unified retrieval model to designate adaptive combination strategies for flexible applicability; 2) Hierarchical Aggregation Transformer aggregates top-down features with Multi-layer Perceptron (MLP) to overcome the limitations of edge information loss in a window-based multi-stage Transformer. Extensive experiments demonstrate the superiority of the proposed CRN over all three fashion-domain datasets. Code is available at github.com/yan9qu/crn.
- 18 July 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 72 (00189456) , 1-16
An Improved Sequence-to-Point Learning for Non-Intrusive Load Monitoring Based on Discrete Wavelet Transform
Nonintrusive load monitoring (NILM) is a process that monitors the aggregated power consumption data of customers measured by a single sensor and decomposes the real-time power consumption of each dedicated device. Recent research has been focused on in-depth learning. However, the expensive cost of training time and the huge model scale are not conducive to the realization of smart meters. It is still challenging to accurately predict the real-time power of high-frequency appliances. This paper uses the discrete wavelet transform (DWT) to preprocess the data, which divides the frequency of the training power data. The processed data will be transferred to the enhanced sequence-to-point network (en-S2P) for training. The en-S2P network is the promotion of a sequence-to-point model (S2P) optimized by the autoencoder and bidirectional long short-term memory (Bi-LSTM) layer. And it is compressed by a combined pruning algorithm after being trained. The proposed model method is tested on the UK-DALE and REDD datasets using db2 and sym5 wavelet transform and compared with the S2P. Extensive experiments show that the proposed model can obtain a more satisfactory decomposition performance of appliances, especially high-frequency. In addition, it has a more lightweight scale and maintains a certain degree of sparsity, which is friendly to electricity meters.
- 18 July 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 72 (00189456) , 1-16
An Improved Sequence-to-Point Learning for Non-Intrusive Load Monitoring Based on Discrete Wavelet Transform
Nonintrusive load monitoring (NILM) is a process that monitors the aggregated power consumption data of customers measured by a single sensor and decomposes the real-time power consumption of each dedicated device. Recent research has been focused on in-depth learning. However, the expensive cost of training time and the huge model scale are not conducive to the realization of smart meters. It is still challenging to accurately predict the real-time power of high-frequency appliances. This paper uses the discrete wavelet transform (DWT) to preprocess the data, which divides the frequency of the training power data. The processed data will be transferred to the enhanced sequence-to-point network (en-S2P) for training. The en-S2P network is the promotion of a sequence-to-point model (S2P) optimized by the autoencoder and bidirectional long short-term memory (Bi-LSTM) layer. And it is compressed by a combined pruning algorithm after being trained. The proposed model method is tested on the UK-DALE and REDD datasets using db2 and sym5 wavelet transform and compared with the S2P. Extensive experiments show that the proposed model can obtain a more satisfactory decomposition performance of appliances, especially high-frequency. In addition, it has a more lightweight scale and maintains a certain degree of sparsity, which is friendly to electricity meters.
- Open Access
- 28 June 2023
- Published by MDPI AG in Algorithms
- Vol. 16 (7) , 317
Audio Anti-Spoofing Based on Audio Feature Fusion
The rapid development of speech synthesis technology has significantly improved the naturalness and human-likeness of synthetic speech. As the technical barriers for speech synthesis are rapidly lowering, the number of illegal activities such as fraud and extortion is increasing, posing a significant threat to authentication systems, such as automatic speaker verification. This paper proposes an end-to-end speech synthesis detection model based on audio feature fusion in response to the constantly evolving synthesis techniques and to improve the accuracy of detecting synthetic speech. The model uses a pre-trained wav2vec2 model to extract features from raw waveforms and utilizes an audio feature fusion module for back-end classification. The audio feature fusion module aims to improve the model accuracy by adequately utilizing the audio features extracted from the front end and fusing the information from timeframes and feature dimensions. Data augmentation techniques are also used to enhance the performance generalization of the model. The model is trained on the training and development sets of the logical access (LA) dataset of the ASVspoof 2019 Challenge, an international standard, and is tested on the logical access (LA) and deep-fake (DF) evaluation datasets of the ASVspoof 2021 Challenge. The equal error rate (EER) on ASVspoof 2021 LA and ASVspoof 2021 DF are 1.18% and 2.62%, respectively, achieving the best results on the DF dataset.
- 28 June 2023
- Published by MDPI AG in Algorithms
- Vol. 16 (7) , 317
Open AccessAudio Anti-Spoofing Based on Audio Feature Fusion
The rapid development of speech synthesis technology has significantly improved the naturalness and human-likeness of synthetic speech. As the technical barriers for speech synthesis are rapidly lowering, the number of illegal activities such as fraud and extortion is increasing, posing a significant threat to authentication systems, such as automatic speaker verification. This paper proposes an end-to-end speech synthesis detection model based on audio feature fusion in response to the constantly evolving synthesis techniques and to improve the accuracy of detecting synthetic speech. The model uses a pre-trained wav2vec2 model to extract features from raw waveforms and utilizes an audio feature fusion module for back-end classification. The audio feature fusion module aims to improve the model accuracy by adequately utilizing the audio features extracted from the front end and fusing the information from timeframes and feature dimensions. Data augmentation techniques are also used to enhance the performance generalization of the model. The model is trained on the training and development sets of the logical access (LA) dataset of the ASVspoof 2019 Challenge, an international standard, and is tested on the logical access (LA) and deep-fake (DF) evaluation datasets of the ASVspoof 2021 Challenge. The equal error rate (EER) on ASVspoof 2021 LA and ASVspoof 2021 DF are 1.18% and 2.62%, respectively, achieving the best results on the DF dataset.
- 18 May 2023
- Published by Springer Nature in The Journal of Supercomputing
- Vol. 79 (16) , 18699-18727
Time-aware neural ordinary differential equations for incomplete time series modeling
Internet of Things realizes the ubiquitous connection of all things, generating countless time-tagged data called time series. However, real-world time series are often plagued with missing values on account of noise or malfunctioning sensors. Existing methods for modeling such incomplete time series typically involve preprocessing steps, such as deletion or missing data imputation using statistical learning or machine learning methods. Unfortunately, these methods unavoidable destroy time information and bring error accumulation to the subsequent model. To this end, this paper introduces a novel continuous neural network architecture, named Time-aware Neural-Ordinary Differential Equations (TN-ODE), for incomplete time data modeling. The proposed method not only supports imputation missing values at arbitrary time points, but also enables multi-step prediction at desired time points. Specifically, TN-ODE employs a time-aware Long Short-Term Memory as an encoder, which effectively learns the posterior distribution from partial observed data. Additionally, the derivative of latent states is parameterized with a fully connected network, thereby enabling continuous-time latent dynamics generation. The proposed TN-ODE model is evaluated on both real-world and synthetic incomplete time-series datasets by conducting data interpolation and extrapolation tasks as well as classification task. Extensive experiments show the TN-ODE model outperforms baseline methods in terms of Mean Square Error for imputation and prediction tasks, as well as accuracy in downstream classification task.
- 18 May 2023
- Published by Springer Nature in The Journal of Supercomputing
- Vol. 79 (16) , 18699-18727
Time-aware neural ordinary differential equations for incomplete time series modeling
Internet of Things realizes the ubiquitous connection of all things, generating countless time-tagged data called time series. However, real-world time series are often plagued with missing values on account of noise or malfunctioning sensors. Existing methods for modeling such incomplete time series typically involve preprocessing steps, such as deletion or missing data imputation using statistical learning or machine learning methods. Unfortunately, these methods unavoidable destroy time information and bring error accumulation to the subsequent model. To this end, this paper introduces a novel continuous neural network architecture, named Time-aware Neural-Ordinary Differential Equations (TN-ODE), for incomplete time data modeling. The proposed method not only supports imputation missing values at arbitrary time points, but also enables multi-step prediction at desired time points. Specifically, TN-ODE employs a time-aware Long Short-Term Memory as an encoder, which effectively learns the posterior distribution from partial observed data. Additionally, the derivative of latent states is parameterized with a fully connected network, thereby enabling continuous-time latent dynamics generation. The proposed TN-ODE model is evaluated on both real-world and synthetic incomplete time-series datasets by conducting data interpolation and extrapolation tasks as well as classification task. Extensive experiments show the TN-ODE model outperforms baseline methods in terms of Mean Square Error for imputation and prediction tasks, as well as accuracy in downstream classification task.
- Open Access
- 15 March 2023
- Published by Springer Nature in Research Square
Water Level Recognition on Resource-Constrained Equipment using Pruned YOLOv5 and Template Matching
This paper proposes a novel water level recognition method for detecting and recognizing water level rulers (WLR) using limited resources of the PX30 hardware development board. Our approach focuses on optimizing network efficiency, recognition speed, and accuracy. Specifically, we use the You Only Look Once Version 5 (YOLOv5) algorithm for object detection, and use the EagleEye algorithm to prune and optimize the model, so as to reduce the number of parameters and computation, and make the model more suitable for running on resource-constrained devices. Though our proposed pruned YOLOv5n@0.6 decrease 2.1% in mean average precision (mAP@0.5) but reduces the number of parameters and computation by 42% and 40%, respectively, compared to the YOLOv5n algorithm. To recognize the water level height, we apply the multi-template matching method to identify the reading number of the water level scale and the equal-ratio method to calculate the water level height. Our method achieves an average RED of 1.26 cm and RER of 5.8% on WLRD with only average 1.16s time cost per image on PX30. The experimental results show that the proposed method has high accuracy and calculation speed and is suitable for computing equipment with limited resources.
- 15 March 2023
- Published by Springer Nature in Research Square
Open AccessWater Level Recognition on Resource-Constrained Equipment using Pruned YOLOv5 and Template Matching
This paper proposes a novel water level recognition method for detecting and recognizing water level rulers (WLR) using limited resources of the PX30 hardware development board. Our approach focuses on optimizing network efficiency, recognition speed, and accuracy. Specifically, we use the You Only Look Once Version 5 (YOLOv5) algorithm for object detection, and use the EagleEye algorithm to prune and optimize the model, so as to reduce the number of parameters and computation, and make the model more suitable for running on resource-constrained devices. Though our proposed pruned YOLOv5n@0.6 decrease 2.1% in mean average precision (mAP@0.5) but reduces the number of parameters and computation by 42% and 40%, respectively, compared to the YOLOv5n algorithm. To recognize the water level height, we apply the multi-template matching method to identify the reading number of the water level scale and the equal-ratio method to calculate the water level height. Our method achieves an average RED of 1.26 cm and RER of 5.8% on WLRD with only average 1.16s time cost per image on PX30. The experimental results show that the proposed method has high accuracy and calculation speed and is suitable for computing equipment with limited resources.
- 6 January 2023
- Published by Springer Nature in The Journal of Supercomputing
- Vol. 79 (8) , 9021-9046
A multi-task learning model for non-intrusive load monitoring based on discrete wavelet transform
- 6 January 2023
- Published by Springer Nature in The Journal of Supercomputing
- Vol. 79 (8) , 9021-9046
A multi-task learning model for non-intrusive load monitoring based on discrete wavelet transform