Jingzhi Zhang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Jingzhi Zhang has a total of 8 co-authors. Here are the top co-authors by number of shared publications:
- Huawei Liu
State Key Laboratory of Heterogeneous Integrated Microsystem, Shanghai Institute of Microsystem and Information Technology, Chinese Academy of Sciences, Shanghai 200050, China
3 shared publications - Baoqing Li
Science and Technology on Microsystem Laboratory, Shanghai Institute of Microsystem and Information Technology, Chinese Academy of Sciences, Shanghai, China
2 shared publications - Baoqing Li
Science and Technology on Microsystem Laboratory, Shanghai Institute of Microsystem and Information Technology, Chinese Academy of Sciences, Shanghai 201800, China
1 shared publication - Hongying Tang
Science and Technology on Microsystem Laboratory, Shanghai Institute of Microsystem and Information Technology, Chinese Academy of Sciences, Shanghai, China
1 shared publication - Jianpo Liu
State Key Laboratory of Heterogeneous Integrated Microsystem, Shanghai Institute of Microsystem and Information Technology, Chinese Academy of Sciences, Shanghai 200050, China
1 shared publication - Jingchang Huang
Shanghai Institute of Micro-System and Information Technology, Chinese Academy of Sciences, Shanghai, China
1 shared publication
Latest Publications
- Open Access
- 1 October 2026
- Published by MDPI AG in Remote Sensing
- Vol. 18 (19) , 3368
Hierarchical Spatial Discrimination for UAV Geo-Localization Under Continuous-Coverage Partial Matching
Cross-view geo-localization enables GNSS-denied UAV navigation by matching a drone image against a geo-tagged satellite database, a key capability for remote sensing. Existing methods perform well on landmark benchmarks but struggle in continuous-coverage scenarios, where a drone image only partially overlaps surrounding tiles and many works sidestep the issue by discarding partial pairs. This setting requires hierarchical spatial discrimination: the model must first separate distant regions and then distinguish nearby tiles using fine-grained spatial layout. We build this hierarchy into three pipeline stages. At the feature level, intermediate Vision Transformer (ViT) layers with Global Average Pooling provide an effective spatially aware representation at no extra parameter cost. At the training level, Dynamic Curriculum Mining (DCM), our primary training contribution, progressively shifts in-batch negative mining from coarse geographic discrimination toward fine-grained visual confusions. At the inference level, the Spatial Overlap Estimation Module (SOEM) estimates drone–satellite overlap from patch-token matching and re-ranks the top candidates. On GTA-UAV cross-area, our method achieves 71.75% Recall@1 and 222.0 m Dis@1; on DenseUAV, it reaches 96.22% Recall@1. In zero-shot cross-dataset evaluation on the real-world UAV-VisLoc dataset, the GTA-UAV-trained model achieves 35.39% Recall@1 without target-domain adaptation, providing additional evidence of transfer to real UAV imagery. Rather than relying on heavy auxiliary modules, the framework deliberately keeps additional architectural overhead minimal: layer truncation reduces backbone inference cost by about 16%, while SOEM introduces only 0.08 M additional parameters.
- 1 October 2026
- Published by MDPI AG in Remote Sensing
- Vol. 18 (19) , 3368
Open AccessHierarchical Spatial Discrimination for UAV Geo-Localization Under Continuous-Coverage Partial Matching
Cross-view geo-localization enables GNSS-denied UAV navigation by matching a drone image against a geo-tagged satellite database, a key capability for remote sensing. Existing methods perform well on landmark benchmarks but struggle in continuous-coverage scenarios, where a drone image only partially overlaps surrounding tiles and many works sidestep the issue by discarding partial pairs. This setting requires hierarchical spatial discrimination: the model must first separate distant regions and then distinguish nearby tiles using fine-grained spatial layout. We build this hierarchy into three pipeline stages. At the feature level, intermediate Vision Transformer (ViT) layers with Global Average Pooling provide an effective spatially aware representation at no extra parameter cost. At the training level, Dynamic Curriculum Mining (DCM), our primary training contribution, progressively shifts in-batch negative mining from coarse geographic discrimination toward fine-grained visual confusions. At the inference level, the Spatial Overlap Estimation Module (SOEM) estimates drone–satellite overlap from patch-token matching and re-ranks the top candidates. On GTA-UAV cross-area, our method achieves 71.75% Recall@1 and 222.0 m Dis@1; on DenseUAV, it reaches 96.22% Recall@1. In zero-shot cross-dataset evaluation on the real-world UAV-VisLoc dataset, the GTA-UAV-trained model achieves 35.39% Recall@1 without target-domain adaptation, providing additional evidence of transfer to real UAV imagery. Rather than relying on heavy auxiliary modules, the framework deliberately keeps additional architectural overhead minimal: layer truncation reduces backbone inference cost by about 16%, while SOEM introduces only 0.08 M additional parameters.
- 15 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 75 (00189456) , 9520515
Hierarchical Cross-Modal Fusion Model for Field Vehicle Classification
Field Vehicle classification is an important task for unattended ground sensor systems. This paper proposes a multimodal fusion-based classification method. Based on field measurement data collected in field environments, each signal sample contains continuous acoustic and seismic signals together with discrete image signals, which leads to information imbalance between the image modality and the acoustic-seismic modalities. To address this issue in multimodal fusion, a Hierarchical Cross-Modal Fusion (HCMF) method is proposed. First, shallow and deep convolutional features are extracted for each modality using a convolutional network. Subsequently, a two-level fusion strategy is performed using the proposed HCMF framework. In the first-level fusion, three intra-modal transformer models are employed to fuse the two convolutional features within each modality, and the outputs of the three modalities are combined into a hybrid modality feature through confidence-weighted fusion. In the second-level fusion, a inter-modal transformer model processes the hybrid modality feature to enable cross-modal information interaction, and the resulting representation is used for classification. Based on a small-scale dataset constructed from field measurement data, a series of experiments are conducted, including model generalization evaluation, ablation studies, modality absence analysis, and algorithm comparisons. Experimental results demonstrate that the proposed model exhibits generalization capability, achieving a macro F1-score of 98.45% with 1.77 million parameters and 81.19 million FLOPs. The proposed method effectively balances classification performance and computational complexity, providing a potential solution for embedded applications in field vehicle classification.
- 15 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 75 (00189456) , 9520515
Hierarchical Cross-Modal Fusion Model for Field Vehicle Classification
Field Vehicle classification is an important task for unattended ground sensor systems. This paper proposes a multimodal fusion-based classification method. Based on field measurement data collected in field environments, each signal sample contains continuous acoustic and seismic signals together with discrete image signals, which leads to information imbalance between the image modality and the acoustic-seismic modalities. To address this issue in multimodal fusion, a Hierarchical Cross-Modal Fusion (HCMF) method is proposed. First, shallow and deep convolutional features are extracted for each modality using a convolutional network. Subsequently, a two-level fusion strategy is performed using the proposed HCMF framework. In the first-level fusion, three intra-modal transformer models are employed to fuse the two convolutional features within each modality, and the outputs of the three modalities are combined into a hybrid modality feature through confidence-weighted fusion. In the second-level fusion, a inter-modal transformer model processes the hybrid modality feature to enable cross-modal information interaction, and the resulting representation is used for classification. Based on a small-scale dataset constructed from field measurement data, a series of experiments are conducted, including model generalization evaluation, ablation studies, modality absence analysis, and algorithm comparisons. Experimental results demonstrate that the proposed model exhibits generalization capability, achieving a macro F1-score of 98.45% with 1.77 million parameters and 81.19 million FLOPs. The proposed method effectively balances classification performance and computational complexity, providing a potential solution for embedded applications in field vehicle classification.
- 20 February 2026
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 37 (7) , 076001
Moving vehicle detection using acoustic and seismic signals with cross-modal fusion model
This paper investigates field moving vehicle detection using field-measured acoustic and seismic signals and proposes a detection algorithm based on cross-modal fusion model. First, the proposed method specifically extracts complementary features from both modalities and designs a transformer-based fusion framework. Second, three key components are further introduced: the modality alignment modules are employed to balance and improve feature representation quality; the bidirectional cross-attention layer is adopted to enable mutual learning and information complementation across modalities; and the gated fusion module is incorporated to ensure effective integration and discriminability of fused features. Finally, we construct a 76 h second-level vehicle detection dataset to evaluate the proposed fusion method. Ablation experiments verify the effectiveness of the important modules and the overall performance of the proposed method. Comparative results against existing detection algorithms and fusion-based methods demonstrate that our approach achieves superior performance, with F1-score of 0.8722, error rate of 0.0720, and AUC of 0.9652. While significantly improving vehicle detection performance, the proposed method maintains a moderate level of computational complexity, making it a promising solution for embedded applications.
- 20 February 2026
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 37 (7) , 076001
Moving vehicle detection using acoustic and seismic signals with cross-modal fusion model
This paper investigates field moving vehicle detection using field-measured acoustic and seismic signals and proposes a detection algorithm based on cross-modal fusion model. First, the proposed method specifically extracts complementary features from both modalities and designs a transformer-based fusion framework. Second, three key components are further introduced: the modality alignment modules are employed to balance and improve feature representation quality; the bidirectional cross-attention layer is adopted to enable mutual learning and information complementation across modalities; and the gated fusion module is incorporated to ensure effective integration and discriminability of fused features. Finally, we construct a 76 h second-level vehicle detection dataset to evaluate the proposed fusion method. Ablation experiments verify the effectiveness of the important modules and the overall performance of the proposed method. Comparative results against existing detection algorithms and fusion-based methods demonstrate that our approach achieves superior performance, with F1-score of 0.8722, error rate of 0.0720, and AUC of 0.9652. While significantly improving vehicle detection performance, the proposed method maintains a moderate level of computational complexity, making it a promising solution for embedded applications.