Weidong Zhang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Weidong Zhang has a total of 19 co-authors. Here are the top co-authors by number of shared publications:
- Jiawen Li
School of Computer Science, Guangdong Polytechnic Normal University, Guangzhou, China
5 shared publications - Jujian Lv
Guangdong Polytechnic Normal University
5 shared publications - Kaihan Lin
Guangdong Polytechnic Normal University
5 shared publications - Rongjun Chen
School of Computer Science, Guangdong Polytechnic Normal University, Guangzhou, China
5 shared publications - Leijun Wang
School of Computer Science, Guangdong Polytechnic Normal University, Guangzhou 510665, China
3 shared publications - Xianglei Hu
School of Computer Science, Guangdong Polytechnic Normal University, Guangzhou, China
3 shared publications
Latest Publications
- 30 September 2026
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 37 (39) , 395701
CMA-YOLO: a lightweight edge‐optimized measurement framework for fish disease detection in smart aquaculture
In smart aquaculture, computer vision-based measurement systems offer a promising alternative to manual inspection for diagnosing fish disease. However, timely measurement remains challenging due to subtle lesion cues, inter‐class similarity, background clutter, and variations in illumination and fish posture. Moreover, existing deep learning methods struggle to balance detection accuracy with practical deployability on resource‐constrained edge devices in the field of measurement. To address these issues, we propose CMA‐YOLO, a lightweight edge-optimized measurement framework that enhances discriminative feature refinement and context‐aware representation learning while keeping the model compact. Specifically, we design two task‐oriented modules: magnitude‐aware linear attention (C2PSA‐MALA) to capture long‐range dependencies, and factorized convolutional attention block (C3k2‐FCAB) to advance multi‐scale context fusion. We also incorporate an information‐preserving downsampling module (ADown) to reduce feature loss during resolution transitions. This framework is evaluated on a self‐built dataset including eight disease categories and one healthy state. Ablation experiments show that CMA‐YOLO achieves mAP@0.5 of 0.8919 and mAP@0.5:0.95 of 0.5967, outperforming the YOLO11n baseline model (0.8416 and 0.5524, respectively). Meanwhile, CMA‐YOLO adds only 0.13 M parameters (from 2.58 M to 2.71 M), maintaining an impressive accuracy‐efficiency trade‐off. Finally, edge deployment tests on an RK3588S device under the Open Neural Network Exchange runtime (37.11 ms, 23.42 FPS) and Rockchip Neural Network-based neural processing unit (NPU) acceleration (16.82 ms NPU inference, 28.04 ms end-to-end latency, 35.66 FPS) demonstrate its practical real‐time inference advantage for fish disease detection. These results indicate that the proposed method provides a deployable measurement solution for detecting fish disease in smart aquaculture systems.
- 30 September 2026
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 37 (39) , 395701
CMA-YOLO: a lightweight edge‐optimized measurement framework for fish disease detection in smart aquaculture
In smart aquaculture, computer vision-based measurement systems offer a promising alternative to manual inspection for diagnosing fish disease. However, timely measurement remains challenging due to subtle lesion cues, inter‐class similarity, background clutter, and variations in illumination and fish posture. Moreover, existing deep learning methods struggle to balance detection accuracy with practical deployability on resource‐constrained edge devices in the field of measurement. To address these issues, we propose CMA‐YOLO, a lightweight edge-optimized measurement framework that enhances discriminative feature refinement and context‐aware representation learning while keeping the model compact. Specifically, we design two task‐oriented modules: magnitude‐aware linear attention (C2PSA‐MALA) to capture long‐range dependencies, and factorized convolutional attention block (C3k2‐FCAB) to advance multi‐scale context fusion. We also incorporate an information‐preserving downsampling module (ADown) to reduce feature loss during resolution transitions. This framework is evaluated on a self‐built dataset including eight disease categories and one healthy state. Ablation experiments show that CMA‐YOLO achieves mAP@0.5 of 0.8919 and mAP@0.5:0.95 of 0.5967, outperforming the YOLO11n baseline model (0.8416 and 0.5524, respectively). Meanwhile, CMA‐YOLO adds only 0.13 M parameters (from 2.58 M to 2.71 M), maintaining an impressive accuracy‐efficiency trade‐off. Finally, edge deployment tests on an RK3588S device under the Open Neural Network Exchange runtime (37.11 ms, 23.42 FPS) and Rockchip Neural Network-based neural processing unit (NPU) acceleration (16.82 ms NPU inference, 28.04 ms end-to-end latency, 35.66 FPS) demonstrate its practical real‐time inference advantage for fish disease detection. These results indicate that the proposed method provides a deployable measurement solution for detecting fish disease in smart aquaculture systems.
- Open Access
- 17 July 2026
- Published by MDPI AG in Veterinary Sciences
- Vol. 13 (7) , 697
BoviFusionNet: A Lightweight Edge-Deployable AI System for Cattle Behavior Recognition in Livestock Monitoring
This study aims to develop a lightweight, edge-deployable artificial intelligence (AI) system for real-time, non-contact recognition of cattle eating, standing, and lying behaviors in farm environments. Automated monitoring of these behaviors in cattle provides fundamental behavioral data for the future development of systems that analyze feeding duration, lying duration, and behavioral rhythms. Nevertheless, practical deployment on farms is hindered by data imbalance, dense animal groupings, scale variation, occlusion, and the need for low-cost edge computing. To address these challenges, we propose BoviFusionNet, a lightweight, edge-deployable AI system. A box balanced augmentation strategy rebalances training instances at the object level without altering the validation or test sets. Built upon YOLO11n, the model integrates three targeted enhancements: information-preserving downsampling (ADown), adaptive bidirectional feature fusion (BiFPN), and local window attention (C2CGA) to improve multi-scale representation and fine-grained behavior discrimination. Experimental results show that BoviFusionNet achieves 0.7851 recall, 0.7763 F1-score, 0.7976 mAP@0.50, and 0.6305 mAP@0.50:0.95, with only 5.4 GFLOPs and a 3.4 MB model size. Compared with the YOLO11n baseline, it improves mAP@0.50:0.95 by 9.92% and reduces the parameter count by 39.8%. After INT8 quantization and deployment on an RK3588S edge device, real-time inference reaches 28.08 frames per second (FPS). Therefore, BoviFusionNet offers an effective accuracy-complexity trade-off for on-farm edge AI applications. By enabling continuous, non-invasive monitoring of health-relevant behaviors, it provides fundamental behavioral data for the future development of veterinary health assessment tools without relying on cloud services or wearable sensors.
- 17 July 2026
- Published by MDPI AG in Veterinary Sciences
- Vol. 13 (7) , 697
Open AccessBoviFusionNet: A Lightweight Edge-Deployable AI System for Cattle Behavior Recognition in Livestock Monitoring
This study aims to develop a lightweight, edge-deployable artificial intelligence (AI) system for real-time, non-contact recognition of cattle eating, standing, and lying behaviors in farm environments. Automated monitoring of these behaviors in cattle provides fundamental behavioral data for the future development of systems that analyze feeding duration, lying duration, and behavioral rhythms. Nevertheless, practical deployment on farms is hindered by data imbalance, dense animal groupings, scale variation, occlusion, and the need for low-cost edge computing. To address these challenges, we propose BoviFusionNet, a lightweight, edge-deployable AI system. A box balanced augmentation strategy rebalances training instances at the object level without altering the validation or test sets. Built upon YOLO11n, the model integrates three targeted enhancements: information-preserving downsampling (ADown), adaptive bidirectional feature fusion (BiFPN), and local window attention (C2CGA) to improve multi-scale representation and fine-grained behavior discrimination. Experimental results show that BoviFusionNet achieves 0.7851 recall, 0.7763 F1-score, 0.7976 mAP@0.50, and 0.6305 mAP@0.50:0.95, with only 5.4 GFLOPs and a 3.4 MB model size. Compared with the YOLO11n baseline, it improves mAP@0.50:0.95 by 9.92% and reduces the parameter count by 39.8%. After INT8 quantization and deployment on an RK3588S edge device, real-time inference reaches 28.08 frames per second (FPS). Therefore, BoviFusionNet offers an effective accuracy-complexity trade-off for on-farm edge AI applications. By enabling continuous, non-invasive monitoring of health-relevant behaviors, it provides fundamental behavioral data for the future development of veterinary health assessment tools without relying on cloud services or wearable sensors.
- 8 July 2026
- Published by Springer Nature in Journal of Food Measurement and Characterization
- Vol. 20 (8) , 16229-16247
WA-YOLO: a lightweight deep learning network model for real-time measurement and characterization of fruit quality
- 8 July 2026
- Published by Springer Nature in Journal of Food Measurement and Characterization
- Vol. 20 (8) , 16229-16247
WA-YOLO: a lightweight deep learning network model for real-time measurement and characterization of fruit quality
- Open Access
An Edge-Ready Lightweight Computer Vision Framework for On-Site Fish Disease Detection in Aquaculture Management
Efficient detection of fish diseases is essential for intelligent health monitoring and timely intervention in aquaculture. However, current computer vision models remain computationally intensive, hindering their deployment on resource-constrained edge devices in aquaculture applications. To this end, this study developed a lightweight detection framework based on an improved You Only Look Once (YOLO), aiming to achieve a favorable balance between detection accuracy and on-site inference efficiency. First, a Dual-Branch Feature-Preserving Downsampling (DFPD) module was proposed to enhance the extraction of valuable disease-related cues with minimal computational overhead. Subsequently, structured pruning was applied to compress the optimized baseline model. Four pruning techniques, including Slim, GroupTaylor, Layer-Adaptive Magnitude-Based Pruning (LAMP), and L1-based, were evaluated under the same conditions. The enhanced baseline model improved precision from 0.864 to 0.908 and mAP@0.5:0.95 from 0.613 to 0.632, while already reducing the Number of Parameters (Params) and Giga Floating-point Operations Per Second (GFLOPs) compared with the original YOLOv8n. Among the pruning techniques, L1-based produced the best overall trade-off, yielding a final model that maintained a F1-score of 0.860 while reducing Params and GFLOPs by 54.7% and 49.4%, respectively, relative to the original detector. Ablation studies further revealed that a moderate FLOPs reduction of approximately 41% to 47% was optimal for preserving diagnostic performance while enhancing compactness. Edge deployment tests on an RK3588S device verified the framework’s practical inference speed advantage. Therefore, this study offers a deployment-friendly computer vision solution for on-site fish disease detection in aquaculture management, particularly suited to real-world scenarios with limited computational resources.
Open AccessAn Edge-Ready Lightweight Computer Vision Framework for On-Site Fish Disease Detection in Aquaculture Management
Efficient detection of fish diseases is essential for intelligent health monitoring and timely intervention in aquaculture. However, current computer vision models remain computationally intensive, hindering their deployment on resource-constrained edge devices in aquaculture applications. To this end, this study developed a lightweight detection framework based on an improved You Only Look Once (YOLO), aiming to achieve a favorable balance between detection accuracy and on-site inference efficiency. First, a Dual-Branch Feature-Preserving Downsampling (DFPD) module was proposed to enhance the extraction of valuable disease-related cues with minimal computational overhead. Subsequently, structured pruning was applied to compress the optimized baseline model. Four pruning techniques, including Slim, GroupTaylor, Layer-Adaptive Magnitude-Based Pruning (LAMP), and L1-based, were evaluated under the same conditions. The enhanced baseline model improved precision from 0.864 to 0.908 and mAP@0.5:0.95 from 0.613 to 0.632, while already reducing the Number of Parameters (Params) and Giga Floating-point Operations Per Second (GFLOPs) compared with the original YOLOv8n. Among the pruning techniques, L1-based produced the best overall trade-off, yielding a final model that maintained a F1-score of 0.860 while reducing Params and GFLOPs by 54.7% and 49.4%, respectively, relative to the original detector. Ablation studies further revealed that a moderate FLOPs reduction of approximately 41% to 47% was optimal for preserving diagnostic performance while enhancing compactness. Edge deployment tests on an RK3588S device verified the framework’s practical inference speed advantage. Therefore, this study offers a deployment-friendly computer vision solution for on-site fish disease detection in aquaculture management, particularly suited to real-world scenarios with limited computational resources.
- Open Access
Multi-Entropy Feature Concatenation for Data-Efficient Cross-Subject Classification of Alzheimer’s Disease and Frontotemporal Dementia from Single-Channel EEG
Alzheimer’s disease (AD) and frontotemporal dementia (FTD) are neurodegenerative disorders where early detection is vital. However, the need for long-term monitoring is incompatible with data-scarce settings, and methods trained on one subject often fail on another due to cross-subject variability. To address these limitations, this study proposes a cross-subject, single-channel electroencephalography (EEG)-based method that uses Multi-Entropy Feature Concatenation (MEFC) to classify AD and FTD. First, single-channel EEG is processed through the Discrete Wavelet Transform (DWT) to extract five rhythms: delta, theta, alpha, beta, and gamma. Subsequently, Permutation Entropy (PE), Singular Spectrum Entropy (SSE), and Sample Entropy (SE) are calculated for each rhythm and concatenated to form a combined MEFC to characterize the non-linear dynamic properties of EEG. Lastly, Dynamic Time Warping (DTW), Pearson Correlation Coefficient (PCC), Wavelet Coherence (WC), and Hilbert Transform Correlation (HTC) are employed to measure the similarity between unknown rhythmic MEFC and those from AD, FTD, and Healthy Control (HC) groups, performing a data-driven classification via similarity measurement. Experimental results on 88 subjects in the AHEPA dataset demonstrate that the beta-rhythm with PCC yields a three-class accuracy of 76.14% using single-channel FP2. In another dataset, the Florida-Based dataset, involving 48 subjects, theta-rhythm with WC achieves a two-class accuracy of 83.33% using FP2. Furthermore, a MATLAB R2023b-based toolbox is developed using the proposed method. Such outcomes are impressive, given the limited data per individual (data-efficient), reliable performance across new subjects (cross-subject), and compatibility with wearable devices (single-channel), providing a novel entropy-based approach for EEG-based applications in biomedical engineering.
Open AccessMulti-Entropy Feature Concatenation for Data-Efficient Cross-Subject Classification of Alzheimer’s Disease and Frontotemporal Dementia from Single-Channel EEG
Alzheimer’s disease (AD) and frontotemporal dementia (FTD) are neurodegenerative disorders where early detection is vital. However, the need for long-term monitoring is incompatible with data-scarce settings, and methods trained on one subject often fail on another due to cross-subject variability. To address these limitations, this study proposes a cross-subject, single-channel electroencephalography (EEG)-based method that uses Multi-Entropy Feature Concatenation (MEFC) to classify AD and FTD. First, single-channel EEG is processed through the Discrete Wavelet Transform (DWT) to extract five rhythms: delta, theta, alpha, beta, and gamma. Subsequently, Permutation Entropy (PE), Singular Spectrum Entropy (SSE), and Sample Entropy (SE) are calculated for each rhythm and concatenated to form a combined MEFC to characterize the non-linear dynamic properties of EEG. Lastly, Dynamic Time Warping (DTW), Pearson Correlation Coefficient (PCC), Wavelet Coherence (WC), and Hilbert Transform Correlation (HTC) are employed to measure the similarity between unknown rhythmic MEFC and those from AD, FTD, and Healthy Control (HC) groups, performing a data-driven classification via similarity measurement. Experimental results on 88 subjects in the AHEPA dataset demonstrate that the beta-rhythm with PCC yields a three-class accuracy of 76.14% using single-channel FP2. In another dataset, the Florida-Based dataset, involving 48 subjects, theta-rhythm with WC achieves a two-class accuracy of 83.33% using FP2. Furthermore, a MATLAB R2023b-based toolbox is developed using the proposed method. Such outcomes are impressive, given the limited data per individual (data-efficient), reliable performance across new subjects (cross-subject), and compatibility with wearable devices (single-channel), providing a novel entropy-based approach for EEG-based applications in biomedical engineering.