Hai Liao
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Hai Liao has a total of 21 co-authors. Here are the top co-authors by number of shared publications:
- Fan Min
School of Computer Science and Software Engineering, Southwest Petroleum University, Chengdu, China
6 shared publications - Song Chen
School of Computer Engineering, Chengdu Technological University, Chengdu, China
6 shared publications - Yan Liang
School of Computer Engineering, Chengdu Technological University, Chengdu, China
5 shared publications Yun Xiao
3 shared publications- Hang Zhou
School of Computer Engineering, Chengdu Technological University, Chengdu, China
2 shared publications - Linyun Xiang
Faculty of Informatics, Eötvös Loránd University, Budapest, Hungary
2 shared publications
Latest Publications
- Open Access
- 15 September 2026
- Published by Elsevier in SSRN Electronic Journal
Detail-Enhanced Multi-Scale Channel-Frequency Fusion for 3D CBCT Segmentation
In complex clinical scenarios, deep learning-based tooth segmentation using cone-beam computed tomography (CBCT) faces two primary challenges. First, tooth loss often leads to the misidentification of adjacent teeth due to the lack of spatial references. Second, low contrast and noise blur tooth boundaries, resulting in the loss of anatomical details. To address these challenges, we propose a Detail-enhanced Multi-scale Channel-frequency Fusion Network (DMCF-Net). To mitigate misidentification of adjacent teeth caused by missing teeth, we introduce structural attention in the frequency domain, leveraging anatomical priors to correct label misclassifications. Concurrently, spatial attention is incorporated in the spatial domain to enhance the model's comprehension of relative tooth positions, thereby facilitating precise identification. To enhance tooth edge perception, we design an edge attention mechanism in the frequency domain to boost contour perception. In the spatial domain, we integrate asymmetric convolutions into a resolution recovery module, ultimately achieving accurate preservation of fine edge details. Validated on the public ToothFairy2 and DentVoxel datasets, DMCF-Net achieves mean Dice coefficients of 91.66\% and 94.38\%, with 95 Hausdorff Distances of 7.80 mm and 4.93 mm, respectively. These results outperform current mainstream methods, demonstrating the effectiveness of our proposed method in resolving clinical segmentation challenges.
- 15 September 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessDetail-Enhanced Multi-Scale Channel-Frequency Fusion for 3D CBCT Segmentation
In complex clinical scenarios, deep learning-based tooth segmentation using cone-beam computed tomography (CBCT) faces two primary challenges. First, tooth loss often leads to the misidentification of adjacent teeth due to the lack of spatial references. Second, low contrast and noise blur tooth boundaries, resulting in the loss of anatomical details. To address these challenges, we propose a Detail-enhanced Multi-scale Channel-frequency Fusion Network (DMCF-Net). To mitigate misidentification of adjacent teeth caused by missing teeth, we introduce structural attention in the frequency domain, leveraging anatomical priors to correct label misclassifications. Concurrently, spatial attention is incorporated in the spatial domain to enhance the model's comprehension of relative tooth positions, thereby facilitating precise identification. To enhance tooth edge perception, we design an edge attention mechanism in the frequency domain to boost contour perception. In the spatial domain, we integrate asymmetric convolutions into a resolution recovery module, ultimately achieving accurate preservation of fine edge details. Validated on the public ToothFairy2 and DentVoxel datasets, DMCF-Net achieves mean Dice coefficients of 91.66\% and 94.38\%, with 95 Hausdorff Distances of 7.80 mm and 4.93 mm, respectively. These results outperform current mainstream methods, demonstrating the effectiveness of our proposed method in resolving clinical segmentation challenges.
- 7 May 2026
- Published by Elsevier in Information Sciences
- Vol. 753
Pre-deployment and fine-tuning strategies of Wi-Fi for irregular coverage
- 7 May 2026
- Published by Elsevier in Information Sciences
- Vol. 753
Pre-deployment and fine-tuning strategies of Wi-Fi for irregular coverage
- Open Access
- 25 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
- Vol. 19 (19391404) , 14077-14094
HAE-Net: Haze-Aware Remote Sensing Object Detection With Local–Global Defogging
Remote sensing object detection under haze extends traditional tasks to extreme scenes with low visibility and image degradation. Its key issues include feature loss during dehazing, inconsistency between the dehazing and detection stages, and limited datasets. In this study, we propose a novel two-stage framework called HAE-Net that integrates a structure-aware dehazing module with a YOLO11s-based object detector. For feature loss during dehazing, our Local and Global Defogging (LGD) network balances global context and local detail restoration to minimize feature loss. Unlike conventional methods, it preserves semantic integrity critical for downstream detection. Regarding inconsistency between the dehazing and detection stages, our Haze-Aware Detector (HAD) employs multi-scale feature fusion and the NWD Loss to optimize dehazed outputs for detection tasks. This enhances feature consistency across stages. For limited datasets, we present a new dataset, HAE-DIOR. It comprises 23,463 high-resolution hazy images and 192,472 annotated targets, spanning 20 object categories. Extensive experiments were conducted on the DIOR dataset, the real hazy datasets Hazy-DIOR and HazyDet, and the synthetic hazy dataset HAE-DIOR.Our method achieves competitive performance across all datasets, with an mAP$_{50}$ of 68.3% on HAE-DIOR, which is 22.7 points higher than YOLO12.
- 25 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
- Vol. 19 (19391404) , 14077-14094
Open AccessHAE-Net: Haze-Aware Remote Sensing Object Detection With Local–Global Defogging
Remote sensing object detection under haze extends traditional tasks to extreme scenes with low visibility and image degradation. Its key issues include feature loss during dehazing, inconsistency between the dehazing and detection stages, and limited datasets. In this study, we propose a novel two-stage framework called HAE-Net that integrates a structure-aware dehazing module with a YOLO11s-based object detector. For feature loss during dehazing, our Local and Global Defogging (LGD) network balances global context and local detail restoration to minimize feature loss. Unlike conventional methods, it preserves semantic integrity critical for downstream detection. Regarding inconsistency between the dehazing and detection stages, our Haze-Aware Detector (HAD) employs multi-scale feature fusion and the NWD Loss to optimize dehazed outputs for detection tasks. This enhances feature consistency across stages. For limited datasets, we present a new dataset, HAE-DIOR. It comprises 23,463 high-resolution hazy images and 192,472 annotated targets, spanning 20 object categories. Extensive experiments were conducted on the DIOR dataset, the real hazy datasets Hazy-DIOR and HazyDet, and the synthetic hazy dataset HAE-DIOR.Our method achieves competitive performance across all datasets, with an mAP$_{50}$ of 68.3% on HAE-DIOR, which is 22.7 points higher than YOLO12.
- Open Access
- 4 March 2026
- Published by Elsevier in SSRN Electronic Journal
LogSNN: Spiking Neural Networks for System Log Anomaly Detection
Log anomaly detection is the most important method for identifying potential system failure. However, current Artificial Neural Networks (ANNs) based methods significantly increase the energy consumption of the system while improving the accuracy of log classification. In this paper, we present a novel SNN framework specifically designed for system log anomaly detection, offering distinct advantages in energy efficiency and temporal information processing. First, we introduce a new LLMs-based Spike Feature Encoding (LSFE) method, which encodes word embeddings from a pre-trained T5-base model into spatiotemporal spike sequences. This approach bridge the gap between high accuracy and low energy consumption in SNNs. Second, we propose an SNN Fine-Tuning (SFT) method that integrates SNN with the T5-base model, allowing supervised fine-tuning to enhance the learning capabilities of the network. This fine-tuning process addresses the accuracy limitations typically encountered in traditional SNNs. Experimental results demonstrate that 1) LogSNN consumes only 23\% of the energy required by UniLog and LogBERT; and 2) LogSNN achieves the highest performance with an F1-score of 0.9951 on the HDFS dataset and 0.9948 on the BGL dataset. This work presents a breakthrough in the application of SNNs to log anomaly detection, positioning SNNs as a competitive solution in resource-constrained environments.
- 4 March 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessLogSNN: Spiking Neural Networks for System Log Anomaly Detection
Log anomaly detection is the most important method for identifying potential system failure. However, current Artificial Neural Networks (ANNs) based methods significantly increase the energy consumption of the system while improving the accuracy of log classification. In this paper, we present a novel SNN framework specifically designed for system log anomaly detection, offering distinct advantages in energy efficiency and temporal information processing. First, we introduce a new LLMs-based Spike Feature Encoding (LSFE) method, which encodes word embeddings from a pre-trained T5-base model into spatiotemporal spike sequences. This approach bridge the gap between high accuracy and low energy consumption in SNNs. Second, we propose an SNN Fine-Tuning (SFT) method that integrates SNN with the T5-base model, allowing supervised fine-tuning to enhance the learning capabilities of the network. This fine-tuning process addresses the accuracy limitations typically encountered in traditional SNNs. Experimental results demonstrate that 1) LogSNN consumes only 23\% of the energy required by UniLog and LogBERT; and 2) LogSNN achieves the highest performance with an F1-score of 0.9951 on the HDFS dataset and 0.9948 on the BGL dataset. This work presents a breakthrough in the application of SNNs to log anomaly detection, positioning SNNs as a competitive solution in resource-constrained environments.
- 15 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 13 (5) , 9375-9395
SNN-Drone: Low-Power Drone-View Object Detection With Integer-Valued Spiking Neural Networks
Drone-view object detection (DroneDet) is a computer vision method for locating objects and predicting their categories in aerial images. Popular methods are often plagued by high energy consumption and severe noise interference in dynamic aerial environments. In this study, we propose a novel integer-valued spiking neural network (SNN)-based approach called SNN-Drone to optimize both energy consumption and performance. Regarding high energy consumption, a new integer-valued encoding (IVE) scheme transforms YOLOv8 features into high-fidelity integer spike trains. The IVE effectively reduces the high energy consumption associated with conventional binary spike encoding. For noise interference, our dual-branch SNN block (DBSB) enhances feature representation by separating spatial and semantic processing, while the normalized Wasserstein distance (NWD) loss refines localization accuracy for small and densely packed targets. These modules significantly improve detection performance. In addition, we construct UAVDT-SNN, the first large-scale neuromorphic dataset for uncrewed aerial vehicle (UAV) object detection with temporally encoded motion and diverse aerial scenes. Extensive experiments on the VisDrone 2019, UAVDT, and UAVDT-SNN datasets demonstrate that SNN-Drone achieves state-of-the-art performance, outperforming both baseline artificial neural networks (ANNs) and advanced YOLOv8–v12 variants. Notably, SNN-Drone-s improves the energy efficiency by $7.7\times $ compared to YOLOv8s. This work is the first to show that SNNs can serve as an energy-efficient yet competitive solution for drone-view object detection, narrowing the performance–energy gap with ANNs. Our code is available at https://github.com/iliaohai/SNN-Drone
- 15 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 13 (5) , 9375-9395
SNN-Drone: Low-Power Drone-View Object Detection With Integer-Valued Spiking Neural Networks
Drone-view object detection (DroneDet) is a computer vision method for locating objects and predicting their categories in aerial images. Popular methods are often plagued by high energy consumption and severe noise interference in dynamic aerial environments. In this study, we propose a novel integer-valued spiking neural network (SNN)-based approach called SNN-Drone to optimize both energy consumption and performance. Regarding high energy consumption, a new integer-valued encoding (IVE) scheme transforms YOLOv8 features into high-fidelity integer spike trains. The IVE effectively reduces the high energy consumption associated with conventional binary spike encoding. For noise interference, our dual-branch SNN block (DBSB) enhances feature representation by separating spatial and semantic processing, while the normalized Wasserstein distance (NWD) loss refines localization accuracy for small and densely packed targets. These modules significantly improve detection performance. In addition, we construct UAVDT-SNN, the first large-scale neuromorphic dataset for uncrewed aerial vehicle (UAV) object detection with temporally encoded motion and diverse aerial scenes. Extensive experiments on the VisDrone 2019, UAVDT, and UAVDT-SNN datasets demonstrate that SNN-Drone achieves state-of-the-art performance, outperforming both baseline artificial neural networks (ANNs) and advanced YOLOv8–v12 variants. Notably, SNN-Drone-s improves the energy efficiency by $7.7\times $ compared to YOLOv8s. This work is the first to show that SNNs can serve as an energy-efficient yet competitive solution for drone-view object detection, narrowing the performance–energy gap with ANNs. Our code is available at https://github.com/iliaohai/SNN-Drone
- 6 December 2025
- Published by Elsevier in Information Sciences
- Vol. 733
Large-scale text-to-SQL generation with adversarial defense
- 6 December 2025
- Published by Elsevier in Information Sciences
- Vol. 733
Large-scale text-to-SQL generation with adversarial defense
- 14 June 2024
- Published by Springer Nature in International Journal of Machine Learning and Cybernetics
- Vol. 16 (1) , 507-521
An effective two-stage training scheme for boundary decision of imbalanced samples
- 14 June 2024
- Published by Springer Nature in International Journal of Machine Learning and Cybernetics
- Vol. 16 (1) , 507-521
An effective two-stage training scheme for boundary decision of imbalanced samples
- Open Access
- 4 March 2024
- Published by Taylor & Francis in Applied Artificial Intelligence
- Vol. 38 (1)
STSG: A Short Text Semantic Graph Model for Similarity Computing Based on Dependency Parsing and Pre-trained Language Models
- 4 March 2024
- Published by Taylor & Francis in Applied Artificial Intelligence
- Vol. 38 (1)
Open AccessSTSG: A Short Text Semantic Graph Model for Similarity Computing Based on Dependency Parsing and Pre-trained Language Models
- Open Access
- 17 December 2023
- Published by Institution of Engineering and Technology (IET) in IET Image Processing
- Vol. 18 (5) , 1178-1188
LIDA‐YOLO: An unsupervised low‐illumination object detection based on domain adaptation
The low‐light environment is integral to everyday activities but poses significant challenges in object detection. Due to the low brightness, noise, and insufficient illumination of the acquired image, the model's object detection performance is reduced. Opposing recent studies mainly developing using supervised learning models, this paper suggests LIDA‐YOLO, an approach for unsupervised adaptation of low‐illumination object detectors. The model improves the YOLOv3 by using normal illumination images as the source domain and low‐illumination images as the target domain and achieves object detection in low‐illumination images through an unsupervised learning strategy. Specifically, a multi‐scale local feature alignment and global feature alignment module are proposed to align the overall attributes of the image and feature biases such as background, scene, and target layout are thus reduced. The experimental results of LIDA‐YOLO on the ExDark dataset achieved the highest performance mAP score of 56.65% compared to several current state‐of‐the‐art unsupervised domain adaptation object detection methods. Compared to I3Net, the performance improvement is 4.04%, and compared to OSHOT, the performance improvement is 6.5%. LIDA‐YOLO achieves a performance improvement of 2.7% compared to the supervised baseline method YOLOv3. Overall, the suggested LIDA‐YOLO model requires fewer samples and presents a stronger generalization ability than previous works.
- 17 December 2023
- Published by Institution of Engineering and Technology (IET) in IET Image Processing
- Vol. 18 (5) , 1178-1188
Open AccessLIDA‐YOLO: An unsupervised low‐illumination object detection based on domain adaptation
The low‐light environment is integral to everyday activities but poses significant challenges in object detection. Due to the low brightness, noise, and insufficient illumination of the acquired image, the model's object detection performance is reduced. Opposing recent studies mainly developing using supervised learning models, this paper suggests LIDA‐YOLO, an approach for unsupervised adaptation of low‐illumination object detectors. The model improves the YOLOv3 by using normal illumination images as the source domain and low‐illumination images as the target domain and achieves object detection in low‐illumination images through an unsupervised learning strategy. Specifically, a multi‐scale local feature alignment and global feature alignment module are proposed to align the overall attributes of the image and feature biases such as background, scene, and target layout are thus reduced. The experimental results of LIDA‐YOLO on the ExDark dataset achieved the highest performance mAP score of 56.65% compared to several current state‐of‐the‐art unsupervised domain adaptation object detection methods. Compared to I3Net, the performance improvement is 4.04%, and compared to OSHOT, the performance improvement is 6.5%. LIDA‐YOLO achieves a performance improvement of 2.7% compared to the supervised baseline method YOLOv3. Overall, the suggested LIDA‐YOLO model requires fewer samples and presents a stronger generalization ability than previous works.
- 17 November 2022
- Published by Taylor & Francis in Applied Artificial Intelligence
- Vol. 36 (1)
BERT-Log: Anomaly Detection for System Logs Based on Pre-trained Language Model
Logs are primary information resource for fault diagnosis and anomaly detection in large-scale computer systems, but it is hard to classify anomalies from system logs. Recent studies focus on extracting semantic information from unstructured log messages and converting it into word vectors. Therefore, LSTM approach is more suitable for time series data. Word2Vec is the up-to-date encoding method, but the order of words in sequences is not taken into account. In this article, we propose BERT-Log, which regards the log sequence as a natural language sequence, use pre-trained language model to learn the semantic representation of normal and anomalous logs, and a fully connected neural network is utilized to fine-tune the BERT model to detect abnormal. It can capture all the semantic information from log sequence including context and position. It has achieved the highest performance among all the methods on HDFS dataset, with an F1-score of 99.3%. We propose a new log feature extractor on BGL dataset to obtain log sequence by sliding window including node ID, window size and step size. BERT-Log approach detects anomalies on BGL dataset with an F1-score of 99.4%. It gives 19% performance improvement compared to LogRobust and 7% performance improvement compared to HitAnomaly.
- 17 November 2022
- Published by Taylor & Francis in Applied Artificial Intelligence
- Vol. 36 (1)
BERT-Log: Anomaly Detection for System Logs Based on Pre-trained Language Model
Logs are primary information resource for fault diagnosis and anomaly detection in large-scale computer systems, but it is hard to classify anomalies from system logs. Recent studies focus on extracting semantic information from unstructured log messages and converting it into word vectors. Therefore, LSTM approach is more suitable for time series data. Word2Vec is the up-to-date encoding method, but the order of words in sequences is not taken into account. In this article, we propose BERT-Log, which regards the log sequence as a natural language sequence, use pre-trained language model to learn the semantic representation of normal and anomalous logs, and a fully connected neural network is utilized to fine-tune the BERT model to detect abnormal. It can capture all the semantic information from log sequence including context and position. It has achieved the highest performance among all the methods on HDFS dataset, with an F1-score of 99.3%. We propose a new log feature extractor on BGL dataset to obtain log sequence by sliding window including node ID, window size and step size. BERT-Log approach detects anomalies on BGL dataset with an F1-score of 99.4%. It gives 19% performance improvement compared to LogRobust and 7% performance improvement compared to HitAnomaly.