Yizhuo Zhang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Yizhuo Zhang has a total of 42 co-authors. Here are the top co-authors by number of shared publications:
- Huiling Yu
Aliyun School of Big Data, School of Software, School of Computer Science and Artificial Intelligence, Changzhou University, Changzhou 213159, China
7 shared publications Dapeng Jiang
6 shared publicationsHuiling Yu
Software Engineering, Department of Computer Science, Changzhou University, Changzhou, 213146, China
5 shared publicationsShen Shi
Changzhou University
5 shared publicationsJinhao Chen
4 shared publications- Shen Shi
School of Computer Science and Artificial Intelligence, Changzhou University, Yanzheng West Avenue, Changzhou 213164, China
4 shared publications
Latest Publications
- Open Access
- 11 September 2026
- Published by MDPI AG in Technologies
- Vol. 14 (9) , 576
CLBD-YOLO: A Lightweight Tea Bud Detection Algorithm
Tea bud detection is a key technology for enabling automated and intelligent tea harvesting. However, accurate detection remains challenging because tea buds are generally small, densely distributed, partially occluded, and visually similar to the surrounding background. In addition, resource-constrained devices impose strict requirements on model size and computational complexity. To address these challenges, a lightweight tea bud detection model, termed CLBD-YOLO, is proposed for resource-constrained tea bud detection applications. Based on YOLOv11n, a C3K2_GD module is designed for the backbone and neck to combine ghost-style lightweight feature generation with input-conditioned dynamic feature modeling. A three-level weighted BiFPN is employed for bidirectional multi-scale feature fusion, while LAE, DAT, and MPDIoU are incorporated for downsampling feature extraction, adaptive spatial representation, and bounding box regression, respectively. Experiments on the predefined TeaBud test subset containing 625 images show that CLBD-YOLO achieves 74.31% precision, 77.07% recall, 85.67% mAP50, and 62.69% mAP50–95. Compared with YOLOv11n, the parameter count is reduced from 2.582 M to 1.426 M and the computational cost from 6.3 G to 5.6 G FLOPs, corresponding to reductions of approximately 44.8% and 11.1%, respectively. For the pre-declared seed-0 run, mAP50 and mAP50–95 decrease by 0.22 and 1.40 percentage points, respectively; when averaged over the three matched random seeds, the corresponding decreases are 0.77 and 1.89 percentage points. Because near-duplicate screening revealed cross-subset augmented siblings in the predefined TeaBud split, these results are interpreted as within-distribution reference values rather than as strictly leakage-free generalization estimates. Direct evaluation on the 591-image external Xinyang Maojian tea bud validation subset using the fixed TeaBud-trained weights without additional fine-tuning or threshold tuning yields 81.2% mAP50 and 62.5% mAP50–95, respectively, for CLBD-YOLO, compared with 80.4% and 61.2% for YOLOv11n. These results show that the compact model remains competitive under this external data distribution. Overall, CLBD-YOLO substantially reduces model complexity while maintaining competitive tea bud detection performance, demonstrating a favorable accuracy–complexity trade-off for resource-constrained applications.
- 11 September 2026
- Published by MDPI AG in Technologies
- Vol. 14 (9) , 576
Open AccessCLBD-YOLO: A Lightweight Tea Bud Detection Algorithm
Tea bud detection is a key technology for enabling automated and intelligent tea harvesting. However, accurate detection remains challenging because tea buds are generally small, densely distributed, partially occluded, and visually similar to the surrounding background. In addition, resource-constrained devices impose strict requirements on model size and computational complexity. To address these challenges, a lightweight tea bud detection model, termed CLBD-YOLO, is proposed for resource-constrained tea bud detection applications. Based on YOLOv11n, a C3K2_GD module is designed for the backbone and neck to combine ghost-style lightweight feature generation with input-conditioned dynamic feature modeling. A three-level weighted BiFPN is employed for bidirectional multi-scale feature fusion, while LAE, DAT, and MPDIoU are incorporated for downsampling feature extraction, adaptive spatial representation, and bounding box regression, respectively. Experiments on the predefined TeaBud test subset containing 625 images show that CLBD-YOLO achieves 74.31% precision, 77.07% recall, 85.67% mAP50, and 62.69% mAP50–95. Compared with YOLOv11n, the parameter count is reduced from 2.582 M to 1.426 M and the computational cost from 6.3 G to 5.6 G FLOPs, corresponding to reductions of approximately 44.8% and 11.1%, respectively. For the pre-declared seed-0 run, mAP50 and mAP50–95 decrease by 0.22 and 1.40 percentage points, respectively; when averaged over the three matched random seeds, the corresponding decreases are 0.77 and 1.89 percentage points. Because near-duplicate screening revealed cross-subset augmented siblings in the predefined TeaBud split, these results are interpreted as within-distribution reference values rather than as strictly leakage-free generalization estimates. Direct evaluation on the 591-image external Xinyang Maojian tea bud validation subset using the fixed TeaBud-trained weights without additional fine-tuning or threshold tuning yields 81.2% mAP50 and 62.5% mAP50–95, respectively, for CLBD-YOLO, compared with 80.4% and 61.2% for YOLOv11n. These results show that the compact model remains competitive under this external data distribution. Overall, CLBD-YOLO substantially reduces model complexity while maintaining competitive tea bud detection performance, demonstrating a favorable accuracy–complexity trade-off for resource-constrained applications.
- Open Access
- 6 September 2026
- Published by MDPI AG in Journal of Imaging
- Vol. 12 (9) , 420
JRTE-GCN: A Graph Convolutional Network with Joint Relation Topology Evolution Modeling for Skeleton-Based Action Recognition
Graph convolutional networks (GCNs) have become a mainstream approach for skeleton-based action recognition by modeling spatial dependencies among human joints. However, existing methods mostly rely on learnable adjacency matrices or average relations over the entire action sequence to construct graph structures. Such designs mainly represent the overall static structure of the skeleton and may be insufficient for characterizing dynamic changes in joint relations during action execution. To address this problem, this paper proposes a graph convolutional network with joint relation topology evolution modeling, termed JRTE-GCN. Specifically, the proposed method first constructs a topology evolution matrix from frame-level joint distance sequences by calculating the temporal standard deviation and the difference between the initial and final stages of an action. This matrix reflects both the fluctuation intensity and stage-wise variation of joint relations. Subsequently, persistent homology is used to extract structural features from the matrix, thereby capturing the dynamic evolution of joint relations. Finally, the extracted evolutionary features are transformed into layer-wise biases and fused with the average topological bias through a learnable residual mechanism. The fused biases are incorporated into the graph convolutional feature extraction process, enabling joint modeling of the overall skeleton structure and dynamic relational changes. Experimental results on NTU RGB+D 60, NTU RGB+D 120, and Northwestern-UCLA show that JRTE-GCN achieves competitive recognition accuracy under multiple evaluation protocols, demonstrating the effectiveness of the proposed method.
- 6 September 2026
- Published by MDPI AG in Journal of Imaging
- Vol. 12 (9) , 420
Open AccessJRTE-GCN: A Graph Convolutional Network with Joint Relation Topology Evolution Modeling for Skeleton-Based Action Recognition
Graph convolutional networks (GCNs) have become a mainstream approach for skeleton-based action recognition by modeling spatial dependencies among human joints. However, existing methods mostly rely on learnable adjacency matrices or average relations over the entire action sequence to construct graph structures. Such designs mainly represent the overall static structure of the skeleton and may be insufficient for characterizing dynamic changes in joint relations during action execution. To address this problem, this paper proposes a graph convolutional network with joint relation topology evolution modeling, termed JRTE-GCN. Specifically, the proposed method first constructs a topology evolution matrix from frame-level joint distance sequences by calculating the temporal standard deviation and the difference between the initial and final stages of an action. This matrix reflects both the fluctuation intensity and stage-wise variation of joint relations. Subsequently, persistent homology is used to extract structural features from the matrix, thereby capturing the dynamic evolution of joint relations. Finally, the extracted evolutionary features are transformed into layer-wise biases and fused with the average topological bias through a learnable residual mechanism. The fused biases are incorporated into the graph convolutional feature extraction process, enabling joint modeling of the overall skeleton structure and dynamic relational changes. Experimental results on NTU RGB+D 60, NTU RGB+D 120, and Northwestern-UCLA show that JRTE-GCN achieves competitive recognition accuracy under multiple evaluation protocols, demonstrating the effectiveness of the proposed method.
- 1 July 2026
- Published by SPIE-Intl Soc Optical Eng in Journal of Applied Remote Sensing
- Vol. 20 (03) , 1-17
LCTCANet: a lightweight CNN-transformer context-aware network for remote sensing super-resolution
Remote sensing imagery is rich in textures and exhibits strong spatial autocorrelation, making both local and global feature extraction essential for super-resolution (SR). Although CNN-based SR methods effectively capture local patterns, their limited receptive field hinders long-range modeling, degrading reconstruction quality. Moreover, their high parameter counts and computational costs restrict deployment on resource-constrained platforms. To address these issues, we propose LCTCANet, a lightweight CNN-Transformer hybrid architecture for remote sensing SR. The network integrates a core feature extraction module composed of a global contextual local block (GCLB) and an edge-structure fusion block (ESFB). GCLB leverages multi-head self-attention to model global dependencies and enhance structural coherence. ESFB, by contrast, focuses on fine-grained local textures through edge-aware attention and structural refinement. Their fusion enables joint learning of global context and local detail, reinforcing high-frequency and structural information. Extensive experiments on NWPU-RESISC45, Draper, and UC Merced datasets demonstrate that LCTCANet achieves competitive SR performance while maintaining low computational cost and parameter efficiency.
- 1 July 2026
- Published by SPIE-Intl Soc Optical Eng in Journal of Applied Remote Sensing
- Vol. 20 (03) , 1-17
LCTCANet: a lightweight CNN-transformer context-aware network for remote sensing super-resolution
Remote sensing imagery is rich in textures and exhibits strong spatial autocorrelation, making both local and global feature extraction essential for super-resolution (SR). Although CNN-based SR methods effectively capture local patterns, their limited receptive field hinders long-range modeling, degrading reconstruction quality. Moreover, their high parameter counts and computational costs restrict deployment on resource-constrained platforms. To address these issues, we propose LCTCANet, a lightweight CNN-Transformer hybrid architecture for remote sensing SR. The network integrates a core feature extraction module composed of a global contextual local block (GCLB) and an edge-structure fusion block (ESFB). GCLB leverages multi-head self-attention to model global dependencies and enhance structural coherence. ESFB, by contrast, focuses on fine-grained local textures through edge-aware attention and structural refinement. Their fusion enables joint learning of global context and local detail, reinforcing high-frequency and structural information. Extensive experiments on NWPU-RESISC45, Draper, and UC Merced datasets demonstrate that LCTCANet achieves competitive SR performance while maintaining low computational cost and parameter efficiency.
- Open Access
- 5 June 2026
- Published by Springer Nature in Journal of Forestry Research
- Vol. 37 (1)
Enhancing EOM with density-adaptive downsampling for individual tree segmentation in density-heterogeneous forest point clouds
Addressing the insufficient segmentation accuracy caused by forest point cloud heterogeneity in existing extreme offset networks for individual tree segmentation, this study proposes a density-adaptive downsampling model with adaptive point cloud density sampling, multi-scale dynamic feature fusion, sample relationship enhancement. First, to resolve key feature points omission from uneven density distribution, the adaptive point cloud density sampling method enhances sparse region sampling probability through local density-weighted scoring. Second, multi-scale dynamic feature fusion module adaptively adjusts neighborhood scales based on local density and integrates cross-scale features. Finally, to clarify tree boundaries, sample relationship enhancement module with external learnable memory units captures global contextual information. For evaluating the method’s accuracy and generalizability across varying forest structures and point cloud densities, experiments are conducted on two low-density coniferous forest plots (U1, U2) in Washington’s Blue Ridge area (USA) and two high-density mixed forest plots (G1, G2) in Bretten (Germany). Results demonstrate average precision, recall, and F1-score of 0.92, 0.92, and 0.92, respectively, outperforming the baseline EOM ( p = 0.91, r = 0.88, F-score = 0.90) with improvements of 1%, 4%, and 2% in precision, recall, and F-score.
- 5 June 2026
- Published by Springer Nature in Journal of Forestry Research
- Vol. 37 (1)
Open AccessEnhancing EOM with density-adaptive downsampling for individual tree segmentation in density-heterogeneous forest point clouds
Addressing the insufficient segmentation accuracy caused by forest point cloud heterogeneity in existing extreme offset networks for individual tree segmentation, this study proposes a density-adaptive downsampling model with adaptive point cloud density sampling, multi-scale dynamic feature fusion, sample relationship enhancement. First, to resolve key feature points omission from uneven density distribution, the adaptive point cloud density sampling method enhances sparse region sampling probability through local density-weighted scoring. Second, multi-scale dynamic feature fusion module adaptively adjusts neighborhood scales based on local density and integrates cross-scale features. Finally, to clarify tree boundaries, sample relationship enhancement module with external learnable memory units captures global contextual information. For evaluating the method’s accuracy and generalizability across varying forest structures and point cloud densities, experiments are conducted on two low-density coniferous forest plots (U1, U2) in Washington’s Blue Ridge area (USA) and two high-density mixed forest plots (G1, G2) in Bretten (Germany). Results demonstrate average precision, recall, and F1-score of 0.92, 0.92, and 0.92, respectively, outperforming the baseline EOM ( p = 0.91, r = 0.88, F-score = 0.90) with improvements of 1%, 4%, and 2% in precision, recall, and F-score.
- Open Access
EHNet: Super-Resolution via Enhanced Dual Convolution and Hybrid-Channel Fusion
Single image super-resolution (SR) is an important part of image processing, which aims to improve the spatial resolution of images. This is a typical ill-posed inverse problem. The main difficulty is that a low-resolution image block usually corresponds to multiple high-resolution image blocks. The existing methods cannot provide enough correlation to determine the unique high-resolution image block, which leads to artifacts and image distortion in the reconstructed image. To address this problem, a method (EHNet) is proposed to achieve super-resolution by using a Hybrid-Channel Fusion Block (HCFB) and an Enhanced Dual-Convolution Block (EDCB). The EDCB effectively enhances the network’s ability to capture image details and textures by combining local and global feature processing. The HCFB strengthens the information interaction between channels by combining channel segmentation with large-kernel convolution, fully explores feature dependencies, and thus optimizes the feature extraction effect. Experimental results show that the super-resolution reconstructed image of EHNet achieves 32.59 dB PSNR and 0.9006 SSIM on the Set5 ×4 SR benchmark, outperforming several state-of-the-art SR methods. In addition, the model exhibits notable improvements in artifact suppression, and the reconstructed image’s subjective visual impact surpasses that of other current techniques.
Open AccessEHNet: Super-Resolution via Enhanced Dual Convolution and Hybrid-Channel Fusion
Single image super-resolution (SR) is an important part of image processing, which aims to improve the spatial resolution of images. This is a typical ill-posed inverse problem. The main difficulty is that a low-resolution image block usually corresponds to multiple high-resolution image blocks. The existing methods cannot provide enough correlation to determine the unique high-resolution image block, which leads to artifacts and image distortion in the reconstructed image. To address this problem, a method (EHNet) is proposed to achieve super-resolution by using a Hybrid-Channel Fusion Block (HCFB) and an Enhanced Dual-Convolution Block (EDCB). The EDCB effectively enhances the network’s ability to capture image details and textures by combining local and global feature processing. The HCFB strengthens the information interaction between channels by combining channel segmentation with large-kernel convolution, fully explores feature dependencies, and thus optimizes the feature extraction effect. Experimental results show that the super-resolution reconstructed image of EHNet achieves 32.59 dB PSNR and 0.9006 SSIM on the Set5 ×4 SR benchmark, outperforming several state-of-the-art SR methods. In addition, the model exhibits notable improvements in artifact suppression, and the reconstructed image’s subjective visual impact surpasses that of other current techniques.
- 9 May 2026
- Published by Springer Nature in Cryptography and Communications
- Vol. 18 (5) , 1-19
Improved parallel linkage construction of constant dimension codes via mixed dimension method
- 9 May 2026
- Published by Springer Nature in Cryptography and Communications
- Vol. 18 (5) , 1-19
Improved parallel linkage construction of constant dimension codes via mixed dimension method
- 1 May 2026
- Published by SPIE-Intl Soc Optical Eng in Journal of Electronic Imaging
- Vol. 35 (03) , 1-23
Mamba-driven approach for multi-crop disease detection
Crop disease detection is crucial for improving crop yield. Current methods mainly focus on single crops, and limited research has addressed challenges such as high inter-crop disease feature similarity, large-scale variations, and low detection accuracy in real field environments. In addition, high local deployment costs remain an issue. To tackle these challenges, we propose a lightweight model, MSN-Mamba-YOLO, for multi-crop and multi-disease detection. First, in the backbone network, an efficient feature extraction module, RG-MMamba, is constructed based on MobileMamba. This module combines wavelet convolution, Mamba, and multi-kernel depthwise separable convolution, effectively alleviating problems caused by high feature similarity and large-scale variations among diseases. Second, to improve detection accuracy in complex field environments, a dual-path parallel scanning mechanism, SN-XSSBlock, is designed in the neck: one path performs global scanning, and the other performs self-structure-aware scanning. By fusing global and structural perception, it achieves precise localization of diseased regions. Finally, in the detection head, a YOLOv11 head integrated with multiple depthwise separable convolutions is adopted, enabling more efficient feature decoupling, reducing computational overhead, and improving overall detection efficiency. On the FieldPlant dataset enhanced with generative-adversarial-network-based augmentation, MSN-Mamba-YOLO achieves an mAP@50–95 of 55.4%, a 1.5% improvement over Mamba-YOLO-T, while reducing parameter count and computational cost by 22% and 30%, respectively. In addition, the model’s cross-crop generalization ability was validated on tea and eggplant disease datasets, demonstrating its efficiency and practicality in multi-crop, multi-disease detection scenarios.
- 1 May 2026
- Published by SPIE-Intl Soc Optical Eng in Journal of Electronic Imaging
- Vol. 35 (03) , 1-23
Mamba-driven approach for multi-crop disease detection
Crop disease detection is crucial for improving crop yield. Current methods mainly focus on single crops, and limited research has addressed challenges such as high inter-crop disease feature similarity, large-scale variations, and low detection accuracy in real field environments. In addition, high local deployment costs remain an issue. To tackle these challenges, we propose a lightweight model, MSN-Mamba-YOLO, for multi-crop and multi-disease detection. First, in the backbone network, an efficient feature extraction module, RG-MMamba, is constructed based on MobileMamba. This module combines wavelet convolution, Mamba, and multi-kernel depthwise separable convolution, effectively alleviating problems caused by high feature similarity and large-scale variations among diseases. Second, to improve detection accuracy in complex field environments, a dual-path parallel scanning mechanism, SN-XSSBlock, is designed in the neck: one path performs global scanning, and the other performs self-structure-aware scanning. By fusing global and structural perception, it achieves precise localization of diseased regions. Finally, in the detection head, a YOLOv11 head integrated with multiple depthwise separable convolutions is adopted, enabling more efficient feature decoupling, reducing computational overhead, and improving overall detection efficiency. On the FieldPlant dataset enhanced with generative-adversarial-network-based augmentation, MSN-Mamba-YOLO achieves an mAP@50–95 of 55.4%, a 1.5% improvement over Mamba-YOLO-T, while reducing parameter count and computational cost by 22% and 30%, respectively. In addition, the model’s cross-crop generalization ability was validated on tea and eggplant disease datasets, demonstrating its efficiency and practicality in multi-crop, multi-disease detection scenarios.
- 27 April 2026
- Published by Elsevier in Digital Signal Processing
- Vol. 179
STAN: A spatio-temporal adaptive graph–transformer network for pedestrian trajectory prediction
- 27 April 2026
- Published by Elsevier in Digital Signal Processing
- Vol. 179
STAN: A spatio-temporal adaptive graph–transformer network for pedestrian trajectory prediction
- 17 February 2026
- Published by Springer Nature in Signal, Image and Video Processing
- Vol. 20 (3) , 100
A novel residual cycle refinement framework for temporal action detection(RCR-TAD)
- 17 February 2026
- Published by Springer Nature in Signal, Image and Video Processing
- Vol. 20 (3) , 100
A novel residual cycle refinement framework for temporal action detection(RCR-TAD)
- 25 August 2025
- Published by Taylor & Francis in International Journal of Remote Sensing
- p. 1-24
Multi-frame remote sensing super-resolution method based on dual-stream feature enhancement
- 25 August 2025
- Published by Taylor & Francis in International Journal of Remote Sensing
- p. 1-24
Multi-frame remote sensing super-resolution method based on dual-stream feature enhancement