Zhen Zheng
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Zhen Zheng has a total of 54 co-authors. Here are the top co-authors by number of shared publications:
- Bingting Zha
School of Mechanical Engineering, Nanjing University of Science and Technology, Nanjing 210018, China
12 shared publications Bingting Zha
8 shared publications- Yu Zhou
School of Mechanical Engineering, Nanjing University of Science and Technology, Nanjing, China
6 shared publications - Guangbo Xu
School of Mechanical Engineering, Nanjing University of Science and Technology, Nanjing 210094, China
5 shared publications Hailu Yuan
5 shared publicationsHe Zhang
5 shared publications
Latest Publications
- Open Access
A Lightweight Feature-Fusion and Small-Target Enhancement Network for Vision-Based UAV Detection
Detecting small unmanned aerial vehicles (UAVs) in ground-to-air imagery is challenging because their weak visual cues must be preserved without imposing excessive computation on resource-constrained platforms. To address the unresolved trade-off between tiny-target representation and deployment efficiency, we propose a Lightweight Feature-Fusion and Small-Target Enhancement Network (LFE-YOLO), a lightweight detector that coordinates partial-channel feature extraction, efficient cross-scale fusion, high-resolution prediction, background-interference suppression, and stable tiny-box regression within a unified architecture. Specifically, C2fFaster and GSConv reduce redundant computation while maintaining multi-scale feature propagation; a P2 high-resolution detection branch and Efficient Multi-scale Attention preserve fine spatial cues and suppress background interference; and Normalized Wasserstein Distance complements Complete Intersection over Union to improve tiny-box localization stability. We also construct Det-UAV by integrating newly collected multi-platform and multi-scene UAV imagery with existing data using scene- and sequence-independent partitioning and duplicate control. Experiments on Det-UAV show that LFE-YOLO improves detection accuracy while reducing parameters and computation relative to YOLOv8s. Zero-shot evaluation on the public DUT Anti-UAV dataset further indicates favorable transferability to an unseen data distribution. TensorRT 8.2.1 deployment experiments on NVIDIA Jetson TX2 show that, under the same evaluation settings, LFE-YOLO achieves higher detection accuracy and inference throughput, lower latency, and a smaller engine size than the comparable-scale YOLOv8s and YOLO11s models. These results support a practical accuracy–efficiency balance for small-UAV detection under constrained resources.
Open AccessA Lightweight Feature-Fusion and Small-Target Enhancement Network for Vision-Based UAV Detection
Detecting small unmanned aerial vehicles (UAVs) in ground-to-air imagery is challenging because their weak visual cues must be preserved without imposing excessive computation on resource-constrained platforms. To address the unresolved trade-off between tiny-target representation and deployment efficiency, we propose a Lightweight Feature-Fusion and Small-Target Enhancement Network (LFE-YOLO), a lightweight detector that coordinates partial-channel feature extraction, efficient cross-scale fusion, high-resolution prediction, background-interference suppression, and stable tiny-box regression within a unified architecture. Specifically, C2fFaster and GSConv reduce redundant computation while maintaining multi-scale feature propagation; a P2 high-resolution detection branch and Efficient Multi-scale Attention preserve fine spatial cues and suppress background interference; and Normalized Wasserstein Distance complements Complete Intersection over Union to improve tiny-box localization stability. We also construct Det-UAV by integrating newly collected multi-platform and multi-scene UAV imagery with existing data using scene- and sequence-independent partitioning and duplicate control. Experiments on Det-UAV show that LFE-YOLO improves detection accuracy while reducing parameters and computation relative to YOLOv8s. Zero-shot evaluation on the public DUT Anti-UAV dataset further indicates favorable transferability to an unseen data distribution. TensorRT 8.2.1 deployment experiments on NVIDIA Jetson TX2 show that, under the same evaluation settings, LFE-YOLO achieves higher detection accuracy and inference throughput, lower latency, and a smaller engine size than the comparable-scale YOLOv8s and YOLO11s models. These results support a practical accuracy–efficiency balance for small-UAV detection under constrained resources.
- 12 June 2026
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 37 (24) , 245206
Fast and accurate graph-based ground filtering of ULS point clouds
Ground filtering is a preliminary and essential step in unmanned aerial vehicle laser scanning (ULS) data processing. Existing methods are mainly developed for airborne laser scanning (ALS) and mobile laser scanning (MLS) point clouds, and exhibit limited applicability to ULS data. This paper proposes a novel graph-based ground filtering method, tailored for ULS point clouds. The proposed method encodes the point cloud into a graph structure and achieves accurate ground point extraction through terrain modeling. To adapt to diverse terrain conditions, we present an innovative solution for labeling connected ground nodes. Specifically, we introduce the spatial connectivity to assess the relationships between neighboring nodes, and design a terrain-adaptive assessment strategy that dynamically adjusts the geometric consistency threshold. Additionally, we devise an improved node weighting scheme with a perturbation index-based correction mechanism for terrain sampling point estimation. This combination captures terrain variation trends and mitigates disturbances caused by near-ground objects, thereby producing a refined global terrain model. Extensive experiments are conducted on real-world ULS point clouds acquired using repetitive and non-repetitive scanning patterns. The results demonstrate that our method achieves outstanding performance, meets real-time requirements, and shows promise for practical ULS applications.
- 12 June 2026
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 37 (24) , 245206
Fast and accurate graph-based ground filtering of ULS point clouds
Ground filtering is a preliminary and essential step in unmanned aerial vehicle laser scanning (ULS) data processing. Existing methods are mainly developed for airborne laser scanning (ALS) and mobile laser scanning (MLS) point clouds, and exhibit limited applicability to ULS data. This paper proposes a novel graph-based ground filtering method, tailored for ULS point clouds. The proposed method encodes the point cloud into a graph structure and achieves accurate ground point extraction through terrain modeling. To adapt to diverse terrain conditions, we present an innovative solution for labeling connected ground nodes. Specifically, we introduce the spatial connectivity to assess the relationships between neighboring nodes, and design a terrain-adaptive assessment strategy that dynamically adjusts the geometric consistency threshold. Additionally, we devise an improved node weighting scheme with a perturbation index-based correction mechanism for terrain sampling point estimation. This combination captures terrain variation trends and mitigates disturbances caused by near-ground objects, thereby producing a refined global terrain model. Extensive experiments are conducted on real-world ULS point clouds acquired using repetitive and non-repetitive scanning patterns. The results demonstrate that our method achieves outstanding performance, meets real-time requirements, and shows promise for practical ULS applications.
- Open Access
MBFTFuse: A Triple-Path Adversarial Network Based on Modality Balancing and Feature-Tracing Compensation for Infrared and Visible Image Fusion
Infrared and visible image fusion aims to integrate complementary information from heterogeneous images captured by different optical sensors based on distinct imaging principles; however, existing methods often exhibit modality bias, leading to weakened targets or the loss of crucial texture details. To address this, we propose MBFTFuse, an adversarial fusion network based on modality balancing and feature tracing, which consists of a triple-path generator and dual discriminators. The architecture employs a generator with a triple-path structure: a central modality-balancing path for deep feature fusion and dual edge feature-tracing paths for modality-specific enhancement. Specifically, a multi-cognitive modality-balancing module is introduced to achieve feature weight equilibrium, while a Feature-Tracing Attention Module self-enhances single-modality features to compensate for information loss in the fusion results. Furthermore, a pixel loss based on intensity histograms is designed to optimize inter-modal balance at the pixel level. Comparative experiments against nine state-of-the-art methods across three public datasets demonstrate that MBFTFuse effectively highlights infrared targets while preserving intricate visible textures. The superior performance of this method in both quantitative metrics and downstream object detection tasks contributes to extending the boundaries of sensor-driven computer vision technologies.
Open AccessMBFTFuse: A Triple-Path Adversarial Network Based on Modality Balancing and Feature-Tracing Compensation for Infrared and Visible Image Fusion
Infrared and visible image fusion aims to integrate complementary information from heterogeneous images captured by different optical sensors based on distinct imaging principles; however, existing methods often exhibit modality bias, leading to weakened targets or the loss of crucial texture details. To address this, we propose MBFTFuse, an adversarial fusion network based on modality balancing and feature tracing, which consists of a triple-path generator and dual discriminators. The architecture employs a generator with a triple-path structure: a central modality-balancing path for deep feature fusion and dual edge feature-tracing paths for modality-specific enhancement. Specifically, a multi-cognitive modality-balancing module is introduced to achieve feature weight equilibrium, while a Feature-Tracing Attention Module self-enhances single-modality features to compensate for information loss in the fusion results. Furthermore, a pixel loss based on intensity histograms is designed to optimize inter-modal balance at the pixel level. Comparative experiments against nine state-of-the-art methods across three public datasets demonstrate that MBFTFuse effectively highlights infrared targets while preserving intricate visible textures. The superior performance of this method in both quantitative metrics and downstream object detection tasks contributes to extending the boundaries of sensor-driven computer vision technologies.
- 26 January 2026
- Published by Optica Publishing Group in Optics Express
- Vol. 34 (3) , 3844-3868
SC-PointLSTM: point cloud target recognition framework for linear array push-broom laser fuze
Laser detection and point cloud processing are key elements of measurement, autonomous driving, and defense technology, particularly for high-precision target recognition and environmental sensing. To meet strict requirements for speed and accuracy in target recognition for laser imaging fuzes, this paper presents a new attempted approach named the SC-PointLSTM framework, which is an innovative, rapid target recognition algorithm designed for linear array push-broom laser fuzes. SC-PointLSTM integrates an enhanced PointNet for efficient sparse point cloud processing with a double-layer long short-term memory (LSTM) for temporal feature fusion, addressing the computational challenges inherent to real-time dynamic environments. This approach enables parallel processing to generate linear 3D point clouds from linear push-broom sensors and effectively fuses previously acquired timestamp sequences, resulting in real-time, accurate target recognition at the end of the push-broom scanning process. The robustness and efficiency of the SC-PointLSTM framework were evaluated on both simulated datasets and real-world targets. The results demonstrated competitive results, achieving an F score exceeding 0.8, with an average recognition time of approximately 32 ms. The SC-PointLSTM framework provides a robust and scalable solution for high-speed target recognition in dynamic environments and real-time laser detection.
- 26 January 2026
- Published by Optica Publishing Group in Optics Express
- Vol. 34 (3) , 3844-3868
SC-PointLSTM: point cloud target recognition framework for linear array push-broom laser fuze
Laser detection and point cloud processing are key elements of measurement, autonomous driving, and defense technology, particularly for high-precision target recognition and environmental sensing. To meet strict requirements for speed and accuracy in target recognition for laser imaging fuzes, this paper presents a new attempted approach named the SC-PointLSTM framework, which is an innovative, rapid target recognition algorithm designed for linear array push-broom laser fuzes. SC-PointLSTM integrates an enhanced PointNet for efficient sparse point cloud processing with a double-layer long short-term memory (LSTM) for temporal feature fusion, addressing the computational challenges inherent to real-time dynamic environments. This approach enables parallel processing to generate linear 3D point clouds from linear push-broom sensors and effectively fuses previously acquired timestamp sequences, resulting in real-time, accurate target recognition at the end of the push-broom scanning process. The robustness and efficiency of the SC-PointLSTM framework were evaluated on both simulated datasets and real-world targets. The results demonstrated competitive results, achieving an F score exceeding 0.8, with an average recognition time of approximately 32 ms. The SC-PointLSTM framework provides a robust and scalable solution for high-speed target recognition in dynamic environments and real-time laser detection.
- 27 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Geoscience and Remote Sensing Letters
- Vol. 23 (1545598X) , 1-5
OVDS-YOLO: An Efficient Detector for Oriented Vehicle Detection in SAR Images
Synthetic aperture radar (SAR) vehicle detection is of critical importance for military reconnaissance and traffic management. Compared to optical images, SAR images exhibit distinctive characteristics. Owing to significant interference from irrelevant objects and limited expressivity of vehicle features, most mainstream detectors based on deep learning face enormous challenges in SAR vehicle detection. Additionally, horizontal bounding boxes (HBBs) are replaced by oriented bounding boxes (OBBs) to suppress background noise and obtain more accurate boundaries of vehicles. However, angle regression remains a challenging problem. To address these issues, we propose an efficient detector for oriented vehicle detection in SAR images, called OVDS-you only look once (YOLO), combining three major improvements. Initially, in order to distinguish between target vehicles and clutter based on backgrounds, we introduce a background information aggregation module (BIAM), which effectively integrates global information and multi-scale surrounding environment information of vehicles. Furthermore, to guide the network to learn salient features of vehicles, we devise a feature selection module (FSM), which focuses on key features at the pixel and channel levels. Finally, to improve the accuracy of angle prediction, we embed a novel joint optimization paradigm (NJOP) into the detection framework, which performs dual-optimization for angles through a combination of Gaussian loss and angle loss. The experimental results on the constructed FARAD vehicle dataset demonstrate that the proposed method achieves excellent performance with low computational overhead.
- 27 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Geoscience and Remote Sensing Letters
- Vol. 23 (1545598X) , 1-5
OVDS-YOLO: An Efficient Detector for Oriented Vehicle Detection in SAR Images
Synthetic aperture radar (SAR) vehicle detection is of critical importance for military reconnaissance and traffic management. Compared to optical images, SAR images exhibit distinctive characteristics. Owing to significant interference from irrelevant objects and limited expressivity of vehicle features, most mainstream detectors based on deep learning face enormous challenges in SAR vehicle detection. Additionally, horizontal bounding boxes (HBBs) are replaced by oriented bounding boxes (OBBs) to suppress background noise and obtain more accurate boundaries of vehicles. However, angle regression remains a challenging problem. To address these issues, we propose an efficient detector for oriented vehicle detection in SAR images, called OVDS-you only look once (YOLO), combining three major improvements. Initially, in order to distinguish between target vehicles and clutter based on backgrounds, we introduce a background information aggregation module (BIAM), which effectively integrates global information and multi-scale surrounding environment information of vehicles. Furthermore, to guide the network to learn salient features of vehicles, we devise a feature selection module (FSM), which focuses on key features at the pixel and channel levels. Finally, to improve the accuracy of angle prediction, we embed a novel joint optimization paradigm (NJOP) into the detection framework, which performs dual-optimization for angles through a combination of Gaussian loss and angle loss. The experimental results on the constructed FARAD vehicle dataset demonstrate that the proposed method achieves excellent performance with low computational overhead.
- Open Access
- 29 August 2025
- Published by MDPI AG in Remote Sensing
- Vol. 17 (17) , 3013
A Novel Laser Detection Method Against Soot Interference Based on Pulse-Width Comparison
Traditional pulsed laser detection systems predominantly utilize threshold-based decision methods that rely on the peak voltage of echo signals for target recognition. However, in environments with soot (smoke and dust) interference, strong backscattering effects lead to severe signal distortion, resulting in false alarms or missed detections, thereby significantly degrading recognition accuracy. To overcome this limitation, through comprehensive simulations and experimental trials, we propose a novel pulse-width comparison method that utilizes the characteristic broadening of echoes induced by soot particles. The method exploits the differences in pulse width between target and soot-induced echoes and achieves effective interference suppression. The target-detection accuracy increases from 71.25% to 93.75% in environments with different concentrations (0–350 mg/), corresponding to a 31.58% relative improvement. This approach leverages pulse-width differences between soot-induced and target echoes to enhance anti-interference capability in dusty environments.- 29 August 2025
- Published by MDPI AG in Remote Sensing
- Vol. 17 (17) , 3013
Open AccessA Novel Laser Detection Method Against Soot Interference Based on Pulse-Width Comparison
Traditional pulsed laser detection systems predominantly utilize threshold-based decision methods that rely on the peak voltage of echo signals for target recognition. However, in environments with soot (smoke and dust) interference, strong backscattering effects lead to severe signal distortion, resulting in false alarms or missed detections, thereby significantly degrading recognition accuracy. To overcome this limitation, through comprehensive simulations and experimental trials, we propose a novel pulse-width comparison method that utilizes the characteristic broadening of echoes induced by soot particles. The method exploits the differences in pulse width between target and soot-induced echoes and achieves effective interference suppression. The target-detection accuracy increases from 71.25% to 93.75% in environments with different concentrations (0–350 mg/), corresponding to a 31.58% relative improvement. This approach leverages pulse-width differences between soot-induced and target echoes to enhance anti-interference capability in dusty environments. - Open Access
Point Cloud Generation Method Based on Dual-Prism Scanning with Multi-Parameter Optimization
This study addresses two critical challenges in biprism-based laser scanning systems: the lack of a comprehensive mathematical framework linking prism parameters to scanning performance, and unresolved theoretical gaps regarding parameter effects on point cloud quality. We propose a multi-parameter optimization method for point cloud generation using dual-prism scanning. By establishing a beam pointing mathematical model, we systematically analyze how prism wedge angles, refractive indices, rotation speed ratios, and placement configurations influence scanning performance, revealing their coupled effects on deflection angles, azimuth control, and coverage. The non-paraxial ray tracing method combined with the Möller–Trumbore algorithm enables efficient point cloud simulation. Experimental results demonstrate that our optimized parameters significantly enhance point cloud density, uniformity, and target feature integrity while overcoming limitations of traditional database construction methods. This work provides both theoretical foundations and practical solutions for high-precision 3D reconstruction in high-speed rendezvous scenarios such as missile-borne laser fuzes, offering advantages in cost-effectiveness and operational reliability.
Open AccessPoint Cloud Generation Method Based on Dual-Prism Scanning with Multi-Parameter Optimization
This study addresses two critical challenges in biprism-based laser scanning systems: the lack of a comprehensive mathematical framework linking prism parameters to scanning performance, and unresolved theoretical gaps regarding parameter effects on point cloud quality. We propose a multi-parameter optimization method for point cloud generation using dual-prism scanning. By establishing a beam pointing mathematical model, we systematically analyze how prism wedge angles, refractive indices, rotation speed ratios, and placement configurations influence scanning performance, revealing their coupled effects on deflection angles, azimuth control, and coverage. The non-paraxial ray tracing method combined with the Möller–Trumbore algorithm enables efficient point cloud simulation. Experimental results demonstrate that our optimized parameters significantly enhance point cloud density, uniformity, and target feature integrity while overcoming limitations of traditional database construction methods. This work provides both theoretical foundations and practical solutions for high-precision 3D reconstruction in high-speed rendezvous scenarios such as missile-borne laser fuzes, offering advantages in cost-effectiveness and operational reliability.
- Open Access
Analysis of Pulsed Laser Target Echo Characteristics in Non-Uniform Smoke Environments
This study establishes a mathematical model for analyzing pulsed laser target echo signals in non-uniform smoke environments, thereby enabling evaluations of the target echo characteristics of laser detection systems under various smoke conditions. A semi-analytical Monte Carlo method for laser reception is developed by integrating the T-matrix scattering phase function rejection method with the characteristics of the non-uniform smoke environment. The effects of the pulse width, smoke concentration, target reflectance, and target distance on the laser echo signal waveform are studied in detail. The results indicate that when the pulse width is small (τ = 5 ns), the echo signal exhibits a dual-peak profile due to the signals from the backscattered smoke particles and the target echo. Moreover, the smoke concentration significantly affects the amplitude of the backscatter signal. When the pulse width is large (τ ≥ 40 ns), the echo signal peak is a combination of both signals, where the amplitude increases with increasing pulse width but decreases with the increasing smoke concentration. Additionally, the target echo signal amplitude is positively correlated with the target reflectance and negatively correlated with the target distance. The accuracy of the proposed model is verified by comparing the simulation results with the experimental data. Overall, this study provides a theoretical foundation for target identification and detection in smoky environments for laser fuze applications and the analysis of laser detection characteristics in smoky environments.
Open AccessAnalysis of Pulsed Laser Target Echo Characteristics in Non-Uniform Smoke Environments
This study establishes a mathematical model for analyzing pulsed laser target echo signals in non-uniform smoke environments, thereby enabling evaluations of the target echo characteristics of laser detection systems under various smoke conditions. A semi-analytical Monte Carlo method for laser reception is developed by integrating the T-matrix scattering phase function rejection method with the characteristics of the non-uniform smoke environment. The effects of the pulse width, smoke concentration, target reflectance, and target distance on the laser echo signal waveform are studied in detail. The results indicate that when the pulse width is small (τ = 5 ns), the echo signal exhibits a dual-peak profile due to the signals from the backscattered smoke particles and the target echo. Moreover, the smoke concentration significantly affects the amplitude of the backscatter signal. When the pulse width is large (τ ≥ 40 ns), the echo signal peak is a combination of both signals, where the amplitude increases with increasing pulse width but decreases with the increasing smoke concentration. Additionally, the target echo signal amplitude is positively correlated with the target reflectance and negatively correlated with the target distance. The accuracy of the proposed model is verified by comparing the simulation results with the experimental data. Overall, this study provides a theoretical foundation for target identification and detection in smoky environments for laser fuze applications and the analysis of laser detection characteristics in smoky environments.
- Open Access
UAV Target Segmentation Based on Depse Unet++ Modeling
To enhance the capabilities of UAV platforms in recognizing and tracking hostile targets on the battlefield, advanced feature extraction and image segmentation are required. In response, the Depse Unet++ model was developed. By introducing Squeeze-and-Excitation, the model’s ability to discriminate camouflaged targets in high-similarity backgrounds is improved; by incorporating a depth-separated convolutional design, the parameters and computational requirements for embedded device applications are significantly reduced; and employing Dropout technique to prevent overfitting with limited sample sizes, thus boosting the model’s adaptability and generalization across environments. Evaluations were conducted using a custom-built dataset to test the model’s segmentation accuracy and performance. The experimental results show that the Depse Unet++ model surpasses existing state-of-the-art models in critical metrics such as Mean Inter-section and Union Ratio, Dice coefficient and pixel accuracy. Additionally, the specific contributions of each technical component to the model’s performance were confirmed through ablation studies, ensuring the method’s validity and reliability. Overall, the Depse Unet++ model offers a more effective tool for target recognition and semantic segmentation for UAVs operating in complex battlefield scenarios.
Open AccessUAV Target Segmentation Based on Depse Unet++ Modeling
To enhance the capabilities of UAV platforms in recognizing and tracking hostile targets on the battlefield, advanced feature extraction and image segmentation are required. In response, the Depse Unet++ model was developed. By introducing Squeeze-and-Excitation, the model’s ability to discriminate camouflaged targets in high-similarity backgrounds is improved; by incorporating a depth-separated convolutional design, the parameters and computational requirements for embedded device applications are significantly reduced; and employing Dropout technique to prevent overfitting with limited sample sizes, thus boosting the model’s adaptability and generalization across environments. Evaluations were conducted using a custom-built dataset to test the model’s segmentation accuracy and performance. The experimental results show that the Depse Unet++ model surpasses existing state-of-the-art models in critical metrics such as Mean Inter-section and Union Ratio, Dice coefficient and pixel accuracy. Additionally, the specific contributions of each technical component to the model’s performance were confirmed through ablation studies, ensuring the method’s validity and reliability. Overall, the Depse Unet++ model offers a more effective tool for target recognition and semantic segmentation for UAVs operating in complex battlefield scenarios.