.png){kind=icon}
.png){kind=icon}
Return
MP-YOLO: a multi-padding strategy framework for enhanced infrared UAV detection
.png){kind=icon}
PRE
AI
Z
L
DOI：10.1117/1.JEI.35.2.023033.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Infrared small target detection presents significant challenges, including limited feature sizes, ambiguity, and the difficulty of distinguishing targets from complex backgrounds. These challenges are particularly pronounced in the detection of infrared unmanned aerial vehicles (UAVs). In addition, convolutional neural network-based object detection models often suffer from information erosion during downsampling due to the use of asymmetric padding, which further increases the difficulty of detecting infrared UAVs. To address these issues, we propose an accurate and efficient object detection framework termed MP-YOLO, which incorporates multiple padding strategies. This framework consists of three innovative plug-and-play modules and one enhanced network layer: the multi-padding module (MPM), the multi-branch fusion module (MBFM), the element-wise multiplication module (EMM), and the small object fusion layer (SOFL). The MPM mitigates the information erosion caused by asymmetric padding while improving the model's capability to extract features from small targets through a larger receptive field and a more uniform distribution of pixel attention. The MBFM employs a multi-branch structure to capture richer contextual information. The EMM transforms the input features into a higher-dimensional nonlinear feature space, which enables the model to learn complex feature mappings. Concurrently, the SOFL concentrates on the integration of feature information specific to small targets. We conducted experiments on two publicly available datasets (Anti-UAV410 and IRSTD-1K), in addition to a self-constructed dataset (thermal UAV dataset). The accuracy of MP-YOLO on these datasets was 0.556, 0.424, and 0.519 (mAP50-95), respectively. The experimental results demonstrate that MP-YOLO outperforms most baseline models as well as several state-of-the-art methods.
Keywords:
infrared small target detection
asymmetric padding
anti-unmanned aerial vehicle
feature mapping
Journal
J
IF:
1
Papers:
109
Citations:
2.7K
Organization
Cited Papers