Zeyuan Hu

Dr. Zeyuan Hu

Information and Communication Engineering, Tongmyong University

Research Focus & Metrics

Academic Career Length

8Years (2019-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
20 June 2024 in Fishes
Verified
Journal Article
Open access
PDF file
YLYinjia Li
Yinjia Li
Zeyuan HuZeyuan Hu
Zeyuan Hu
艺张艺燨 张
艺燨 张
JLJihang Liu
Jihang Liu
WTWan Tu
Wan Tu
HYHong Yu
Hong Yu
Accurately detecting and counting abnormal fish behaviors in aquaculture is essential. Timely detection allows farmers to take swift action to protect fish health and prevent economic losses. This paper proposes an enhanced high-precision detection algorithm based on YOLOv9, named DDEYOLOv9, to facilitate the detection and counting of abnormal fish behavior in industrial aquaculture environments. To address the lack of publicly available datasets on abnormal behavior in fish, we created the “Abnormal Behavior Dataset of Takifugu rubripes”, which includes five categories of fish behaviors. The detection algorithm was further enhanced in several key aspects. Firstly, the DRNELAN4 feature extraction module was introduced to replace the original RepNCSPELAN4 module. This change improves the model’s detection accuracy for high-density and occluded fish in complex water environments while reducing the computational cost. Secondly, the proposed DCNv4-Dyhead detection head enhances the model’s multi-scale feature learning capability, effectively recognizes various abnormal fish behaviors, and improves the computational speed. Lastly, to address the issue of sample imbalance in the abnormal fish behavior dataset, we propose EMA-SlideLoss, which enhances the model’s focus on hard samples, thereby improving the model’s robustness. The experimental results demonstrate that the DDEYOLOv9 model achieves high Precision, Recall, and mean Average Precision (mAP) on the “Abnormal Behavior Dataset of Takifugu rubripes”, with values of 91.7%, 90.4%, and 94.1%, respectively. Compared to the YOLOv9 model, these metrics are improved by 5.4%, 5.5%, and 5.4%, respectively. The model also achieves a running speed of 119 frames per second (FPS), which is 45 FPS faster than YOLOv9. Experimental results show that the DDEYOLOv9 algorithm can accurately and efficiently identify and quantify abnormal fish behaviors in specific complex environments.

24 September 2020 in Symmetry
Verified
Journal Article
Open access
PDF file
Zeyuan HuZeyuan Hu
Zeyuan Hu
ELEung-Joo Lee
Eung-Joo Lee
Traditional convolution neural networks have achieved great success in human action recognition. However, it is challenging to establish effective associations between different human bone nodes to capture detailed information. In this paper, we propose a dual attention-guided multiscale dynamic aggregate graph convolution neural network (DAG-GCN) for skeleton-based human action recognition. Our goal is to explore the best correlation and determine high-level semantic features. First, a multiscale dynamic aggregate GCN module is used to capture important semantic information and to establish dependence relationships for different bone nodes. Second, the higher level semantic feature is further refined, and the semantic relevance is emphasized through a dual attention guidance module. In addition, we exploit the relationship of joints hierarchically and the spatial temporal correlations through two modules. Experiments with the DAG-GCN method result in good performance on the NTU-60-RGB+D and NTU-120-RGB+D datasets. The accuracy is 95.76% and 90.01%, respectively, for the cross (X)-View and X-Subon the NTU60dataset.

1 November 2019
Verified
Conference Paper
Closed access
Zeyuan HuZeyuan Hu
Zeyuan Hu
ELEung-Joo Lee
Eung-Joo Lee
In recent years, deep Convolutional neural networks(CNNs) have made fantastic progress in static image recognition, but the ability to model motion information on behavioral video is weak. Therefore, our paper put forward a new time transition layer that models variable temporal convolution kernel depths. We embed this new Hybrid Model in our proposed 3D CNN. We extend the DenseNet architecture with 3D filters and pooling kernels. It will take time as training a 3D convolutional neural network requires a large number of tagged data sets to start training from the input. Therefore, the focus of this paper is on simple and effective technique of passing 2D convolutional neural network pre-trained data to a randomly initialized 3D convolutional neural network for stable weight initialization, where we can still achieve our experimental results by appropriately reducing the number of 3D convolutional neural network training samples. Experiments show that the network can make a more accurate classification of behavioral video, identify it in the UCF-101 database, and compare it with other classical algorithms that have appeared in recent years. The results reflect the superiority of the algorithm.

1 January 2024 in Aquaculture Research
Verified
Journal Article
Open access
JLJihang Liu
Jihang Liu
Zeyuan HuZeyuan Hu
Zeyuan Hu
YZYixi Zhang
Yixi Zhang
YLYinjia Li
Yinjia Li
JYJinsong Yang
Jinsong Yang
HYHong Yu
Hong Yu
To address the issues of accurately identifying and tracking individual fish abnormal behaviors and poor adaptability in the aquaculture field, this paper proposes a Mask2former model combined with a bidirectional routing attention mechanism (BiFormer) and a multiscale dilated attention (MSDA) module for fish abnormal behavior recognition and segmentation. To compensate for the lack of publicly available datasets on fish abnormal behavior, we created the “FISH_segmentation_2023” abnormal behavior dataset, which includes four types of fish behaviors. First, by introducing the BiFormer attention mechanism, the model can better capture critical temporal and spatial information in image sequences, significantly enhancing feature representation. Second, after processing the feature maps with the pixel decoder, the MSDA module is introduced to perform multiscale fusion on these features. The fused features are then passed to the transformer decoder, further enhancing the model’s ability to recognize fish abnormal behaviors. Finally, to further improve model performance and address class imbalance issues in the dataset, we designed a composite loss function combining focal loss and dice loss (FD loss). This loss function can balance the influence of easy and difficult‐to‐classify samples while optimizing segmentation performance, thereby improving the model’s recognition accuracy and mean intersection over union (mIoU) metrics. Experimental results show that the BiFormer multiscale dilated attention FD loss (BMF)‐Mask2former model exhibits high performance, achieving average intersection over union (IoU), accuracy, and recall values of 92.33%, 95.63%, and 94.82%, respectively, on the self‐built FISH_segmentation_2023 dataset, representing improvements of 6.10%, 4.50%, and 5.09%, respectively, compared to the Mask2former model. The study demonstrates that the proposed model can accurately capture both local and contextual features of fish abnormal behaviors through multiscale fusion methods, resulting in high‐quality segmentation outcomes.

23 February 2026 in arXiv
Verified
Preprint
Open access
PDF file
JYJinsong Yang
Jinsong Yang
Zeyuan HuZeyuan Hu
Zeyuan Hu
YLYichen Li
Yichen Li
HYHong Yu
Hong Yu
Underwater fish detection (UFD) is a core capability for smart aquaculture and marine ecological monitoring. While recent detectors improve accuracy by stacking feature extractors or introducing heavy attention modules, they often incur substantial computational overhead and, more importantly, neglect the physics that fundamentally limits UFD: wavelength-dependent absorption and turbidity-induced scattering significantly degrade contrast, blur fine structures, and introduce backscattering noise, leading to unreliable localization and recognition. To address these challenges, we propose FinSight-Net, an efficient and physics-aware detection framework tailored for complex aquaculture environments. FinSight-Net introduces a Multi-Scale Decoupled Dual-Stream Processing (MS-DDSP) bottleneck that explicitly targets frequency-specific information loss via heterogeneous convolutional branches, suppressing backscattering artifacts while compensating distorted biological cues through scale-aware and channel-weighted pathways. We further design an Efficient Path Aggregation FPN (EPA-FPN) as a detail-filling mechanism: it restores high-frequency spatial information typically attenuated in deep layers by establishing long-range skip connections and pruning redundant fusion routes, enabling robust detection of non-rigid fish targets under severe blur and turbidity. Extensive experiments on DeepFish, AquaFishSet, and our challenging UW-BlurredFish benchmark demonstrate that FinSight-Net achieves state-of-the-art performance. In particular, on UW-BlurredFish, FinSight-Net reaches 92.8% mAP, outperforming YOLOv11s by 4.8% while reducing parameters by 29.0%, providing a strong and lightweight solution for real-time automated monitoring in smart aquaculture.

17 July 2026 in Fishes
Verified
Journal Article
Open access
PDF file
JYJinsong Yang
Jinsong Yang
HTHongxi Tao
Hongxi Tao
Zeyuan HuZeyuan Hu
Zeyuan Hu
YWYue Wang
Yue Wang
HYHong Yu
Hong Yu
Underwater fish detection (UFD) is pivotal for smart aquaculture and marine ecological monitoring. While recent detectors improve accuracy using stacked feature extractors or heavy attention modules, they often incur substantial computational overhead and, more importantly, neglect the underwater optical degradation that fundamentally limits UFD: wavelength-dependent absorption and turbidity-induced scattering significantly degrade contrast, blur fine structures, and introduce back-scattering noise, leading to unreliable localization and recognition. To address these challenges, we propose FinSight-Net, an efficient and prior-guided detection framework tailored to complex aquaculture environments. FinSight-Net introduces a Prior-guided Optical Decoupling (PIOD) bottleneck for degradation-aware underwater feature representation. PIOD decomposes features into four degradation-aware branches—forward-scattering blur, back-scattering veil, spectral attenuation, and high-frequency structural decay—and supervises their degradation response maps using image-derived pseudo priors. We further design a Sparsity-constrained High-frequency Adaptive Routing Pyramid (SHARP-FPN), which uses gradient-based salience responses and sparse cross-scale fusion to preserve boundary-sensitive cues and reduce redundant skip contributions, improving fish localization under severe blur and turbidity. Extensive experiments on DeepFish, AquaFishSet, and our challenging UW-BlurredFish benchmark demonstrate that FinSight-Net achieves state-of-the-art performance. In particular, on UW-BlurredFish, FinSight-Net achieves a state-of-the-art 92.8% mAP50, outperforming YOLOv11s, YOLOv12s, and YOLO26s by 4.8%, 2.0%, and 1.8%, respectively, while reducing the parameter count by 28.7%, 28.0%, and 29.5%, providing a strong and lightweight solution for real-time automated monitoring in smart aquaculture.