Jianhua Huang

Dr. Jianhua Huang


Research Focus & Metrics

Academic Career Length

4Years (2023-Present)

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

Representative Publications

View All Publications
11 November 2025 in Complex & Intelligent Systems
Verified
Journal Article
Open access
PDF file
JLJian Luo
Jian Luo
KCKaixuan Chang
Kaixuan Chang
JHJianhua Huang
Jianhua Huang
XSXiyan Sun
Xiyan Sun
JHJianhua Huang
Jianhua Huang
In recent years, with the rapid development and deep integration of UAV technology and deep learning, UAV aerial photography small object detection has been widely used in urban traffic monitoring, disaster relief, military, and other fields. However, the currently popular deep learning methods encounter several challenges when directly applied to small object detection in UAV aerial images. These include difficulties in detecting dense objects, loss of fine-grained information, and inadequate feature representation of small objects. Additionally, there are issues such as mutual dilution and semantic conflicts between adjacent features during multi-level feature fusion. To address these issues, this paper proposes the fine-grained feature preservation and multi-scale feature pyramid balancing network (FMFN-YOLO), which is based on the YOLO architecture. Firstly, we investigated the impact of various scale detection heads on model performance. Based on our findings, we readjusted the network structure, thereby enhancing the model’s ability to detect dense objects. Secondly, we proposed three innovative plug-and-play modules: the fine-grained preservation feature aggregation module (FPFA), the multi-scale feature enhancement module (MFEM), and the feature pyramid balancing module (FPBM). These three modules are respectively used to enhance the network’s capabilities in fine-grained information retention, multi-scale feature extraction, and multi-level feature fusion, while effectively suppressing interference from complex backgrounds. We demonstrate the effectiveness of our method on the VisDrone2019 and AI-TOD datasets, where FMFN-YOLO achieves a mAP50 of 44.5 and 43.6%, respectively. This represents an improvement of 7.6 and 6.4% over the baseline, respectively. Similarly, it offers significant advantages over state-of-the-art methods.

1 July 2025 in Complex & Intelligent Systems
Verified
Journal Article
Open access
PDF file
KCKaixuan Chang
Kaixuan Chang
JHJianhua Huang
Jianhua Huang
XSXiyan Sun
Xiyan Sun
JHJianhua Huang
Jianhua Huang
SBShitao Bao
Shitao Bao
HHHuansheng Huang
Huansheng Huang
Currently, single-sensor data is frequently utilized in technologies such as object detection. However, in certain scenarios, some sensors may experience failure or information loss, significantly impacting model performance. Given the notable complementarity between infrared and visible images at the information level, it is helpful to improve the robustness and reliability of the model by fusing them and applying them to object detection and other technologies. Nevertheless, most prevalent infrared and visible image fusion methods focus on exploring invariant features across multimodal images, somewhat neglecting the inherent characteristics of the images themselves, leading to issues like structural blurriness and unclear detailed textures in the fused images, which fail to meet application demands. To overcome this challenge, this paper proposes an Infrared and visible image fusion network based on multistage progressive injection, termed MPIFusion. To effectively leverage the inherent characteristics and complementarity of images, and address the issues of structural blurriness and unclear detailed textures in fused images, we introduce a Dual-channel Shallow Detail Fusion Module (DC-SDFM) and a Deep Feature Fusion Block (DFFB). These modules first enhance the original features and then fuse the hierarchical features of infrared and visible images with the aid of an attention mechanism module. Furthermore, we construct a progressive injection layer based on the Information Fusion Module (IFM), integrating the fused features within the same framework to generate high-quality infrared and visible fused images. Extensive experiments demonstrate that our MPIFusion outperforms 15 existing fusion methods in terms of performance. The generated fused images not only highlight global and local detail features but also exhibit higher clarity and contrast. Finally, we apply fusion methods to object detection scenarios, and the results show that MPIFusion exhibits significant superiority in such scenarios, providing more robust and reliable image support. The source code is available at https://github.com/Kaixuan-Chang/MPIFusion.

15 September 2023
Verified
Conference Paper
Open access
ZYZhengyao Yu
Zhengyao Yu
JHJianhua Huang
Jianhua Huang
Yuanfa JiYuanfa Ji
Yuanfa Ji
MLMingming Luo
Mingming Luo
YWYixuan Wan
Yixuan Wan
Traditional change detection methods suffer from limited precision and inadequate feature discriminability in images. Although deep learning has considerably advanced the building change detection field, challenges persist due to the diverse change types in remote sensing imagery across distinct epochs, leading to complex backdrops and arduous feature extraction. To address this issue, we propose a novel approach that combines convolutional neural networks (CNNs) with Swin Transformers to further enhance the accuracy of building change detection in remote sensing images. Firstly, CNNs are employed for preliminary feature extraction from input images. Subsequently, the Transformer's self-attention mechanism is utilized to obtain more comprehensive semantic information about the building structures within the feature maps. Lastly, leveraging the multi-level feature extraction advantages of Swin Transformers, we construct a network capable of fusing and extracting building change features in remote sensing images. By introducing a differential feature fusion module during the decoding phase, the model's receptive field in multi-scale feature fusion processes is enhanced, along with the reinforcement of local details and improvement of segmentation precision. Experimental comparisons on the WUH-CD dataset reveal that, compared to the classic change detection network FC-EF, our proposed method has yielded improvements of 1.7%, 23.7%, 18.7%, 13.5%, and 27.5% in overall accuracy, precision, F1, recall rate, and IoU, respectively.