Chenhui Fan
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Research Interests
Chenhui Fan has a total of 1 co-authors. Here are the top co-authors by number of shared publications:
- Zhengyu Li
Shenyang Jianzhu University
1 shared publication
Latest Publications
- 23 June 2026
- Published by IOP Publishing in Engineering Research Express
- Vol. 8 (13) , 135239
EST-YOLO: An Efficient Mamba-Enhanced Detector for UAV Remote Sensing Object Detection
UAV target detection refers to the use of computer vision algorithms to automatically identify specific objects in aerial images captured by UAV cameras. Due to the high-altitude viewpoint, these images often contain numerous small targets, which are easily missed. Existing methods are typically based on either CNN architectures or attention mechanisms. The former has insufficient global perception capabilities, making it difficult to identify small targets, while the latter requires excessive memory consumption, resulting in slow inference speed. Moreover, region-partitioned attention mechanisms often fail to effectively exchange information across different blocks, leading to inadequate extraction of small-object features. To address these issues, we propose EST-YOLO (Efficient Small Target YOLO), a network with frequency--spatial perception capabilities. First, we design a lightweight backbone, named SGGNet, which introduces a spectral-guided ghost feature generation mechanism to enable frequency-aware and low-cost feature extraction while reducing model parameters. Second, we design PFSA-FMamba (Progressive Frequency-Spatial Attention with Frequency Mamba), which first divides the feature map into blocks to ensure real-time performance, and then performs Mamba-based frequency modeling together with spatial attention to strengthen the focus on small-object regions and improve detection accuracy. Finally, we introduce a Small Object Detection Head (SODP2), which replaces the deep, low-resolution branch with a shallow, high-resolution one to preserve fine-grained spatial details of small targets while reducing redundant computation. With 49.62% fewer parameters, EST-YOLO achieves 9.98% and 5.29% higher mAP50 on VisDrone and UAVDT, respectively, compared with YOLOv12.
- 23 June 2026
- Published by IOP Publishing in Engineering Research Express
- Vol. 8 (13) , 135239
EST-YOLO: An Efficient Mamba-Enhanced Detector for UAV Remote Sensing Object Detection
UAV target detection refers to the use of computer vision algorithms to automatically identify specific objects in aerial images captured by UAV cameras. Due to the high-altitude viewpoint, these images often contain numerous small targets, which are easily missed. Existing methods are typically based on either CNN architectures or attention mechanisms. The former has insufficient global perception capabilities, making it difficult to identify small targets, while the latter requires excessive memory consumption, resulting in slow inference speed. Moreover, region-partitioned attention mechanisms often fail to effectively exchange information across different blocks, leading to inadequate extraction of small-object features. To address these issues, we propose EST-YOLO (Efficient Small Target YOLO), a network with frequency--spatial perception capabilities. First, we design a lightweight backbone, named SGGNet, which introduces a spectral-guided ghost feature generation mechanism to enable frequency-aware and low-cost feature extraction while reducing model parameters. Second, we design PFSA-FMamba (Progressive Frequency-Spatial Attention with Frequency Mamba), which first divides the feature map into blocks to ensure real-time performance, and then performs Mamba-based frequency modeling together with spatial attention to strengthen the focus on small-object regions and improve detection accuracy. Finally, we introduce a Small Object Detection Head (SODP2), which replaces the deep, low-resolution branch with a shallow, high-resolution one to preserve fine-grained spatial details of small targets while reducing redundant computation. With 49.62% fewer parameters, EST-YOLO achieves 9.98% and 5.29% higher mAP50 on VisDrone and UAVDT, respectively, compared with YOLOv12.