Gaofeng Zhu

Gaofeng Zhu

Auto-generated profile, based on publicly available information

Gaofeng Zhu

School of Rail Transit, Shandong Jiaotong University, Jinan 250300, China

Research Focus & Metrics

Academic Career Length

3Years (2024-Present)

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

Representative Publications

View All Publications
3 April 2024 in IEEE Journal of Radio Frequency Identification
Verified
Journal Article
Closed access
GZGaofeng Zhu
Gaofeng Zhu
ZWZhixue Wang
Zhixue Wang
Fenghua ZhuFenghua Zhu
Fenghua Zhu
Gang XiongGang Xiong
Gang Xiong
ZLzheng Li
zheng Li
GSGang Shen
Gang Shen
Drone detection plays a key role in various fields, but from the perspective of drones, factors such as the size of the target, interference from different backgrounds, and lighting affect the detection effect, which can easily lead to missed detections and false detections. To address this problem, this paper proposes a small target detection algorithm. First, the hybrid control of attention mechanism and a convolutional module (HCAC) are used to effectively extract contextual details of targets of different scales, directions, and shapes, while relative position encoding is used to associate targets with position information. Secondly, in view of the small size characteristics of small targets, a high-resolution detection branch is introduced, the large target detection head and its redundant network layers are pruned, and a multi-level weighted feature fusion network (MWFN) is used for multi-dimensional fusion. Finally, the WIoU loss is used as a bounding box regression loss, combined with a dynamic non-monotonic focusing mechanism, to evaluate the quality of anchor boxes so that the detector can handle anchor boxes of different qualities, thus improving the overall performance. Experiments were conducted on the UAV aerial photography data set VisDrone2019. The results showed that the accuracy of P increased by 9.0% and MAP by 9.8%, with higher detection results.

8 January 2025 in Electronics
Verified
Journal Article
Open access
PDF file
GZGaofeng Zhu
Gaofeng Zhu
Fenghua ZhuFenghua Zhu
Fenghua Zhu
ZWZhixue Wang
Zhixue Wang
GZGaofeng Zhu
Gaofeng Zhu
ZLzheng Li
zheng Li
Unmanned aerial vehicles (UAVs) integrated with computer vision technology have emerged as an effective method for information acquisition in various applications. However, due to the small proportion of target pixels and susceptibility to background interference in multi-angle UAV imaging, missed detections and false results frequently occur. To address this issue, a small target detection algorithm, EDANet, is proposed based on YOLOv8. First, the backbone network is replaced by EfficientNet, which can dynamically explore the network size and the image resolution using a scaling factor. Second, the EC2f feature extraction module is designed to achieve unique coding in different directions through parallel branches. The position information is effectively embedded in the channel attention to enhance the spatial representation ability of features. To mitigate the low utilization of small target pixels, we introduce the DTADH detection module, which facilitates feature fusion via a feature-sharing interactive network. Simultaneously, a task alignment predictor assigns classification and localization tasks. In this way, not only is feature utilization optimized, but also the number of parameters is reduced. Finally, leveraging logic and feature knowledge distillation, we employ binary probability mapping of soft labels and a soft label weighting strategy to enhance the algorithm’s learning capabilities in target classification and localization. Experimental validation on the UAV aerial dataset VisDrone2019 demonstrates that EDANet outperforms existing methods, reducing GFLOPs by 39.3% and improving Map by 4.6%.