LMANet: A Lighter and More Accurate Multiobject Detection Network for UAV Remote Sensing Imagery

• Published in: IEEE geoscience and remote sensing letters Vol. 21; pp. 1 - 5
• Main Authors: Fu, Qingwei, Zheng, Qianying, Yu, Fan
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Autonomous aerial vehicles; Balances (scales); Complexity; Computation; Computational modeling; Computer applications; Convolution; Efficient feature extraction; Environmental information; Feature extraction; Head; Image reconstruction; lightweight; model output reconstruction (MOR); multiobject detection; Object recognition; Regression analysis; Regression models; Remote sensing; Task analysis; unmanned aerial vehicle (UAV); Unmanned aerial vehicles
• ISSN: 1545-598X; 1558-0571
• DOI: 10.1109/LGRS.2024.3432329

Object detection using unmanned aerial vehicle (UAV) remote sensing images is a challenging task due to varying object scales, dense distribution, and the predominance of small object. Directly using a generalized object detector that has not been specially designed makes it difficult to balance accuracy and model complexity. To address this challenge, we propose a lighter and more accurate network (LMANet) model. First, a more effective loss function called IMIoU has been developed by combining the concepts of minimum point distance bounding box regression-based loss with auxiliary edge-assisted regression. Second, the model output reconstruction (MOR) was used to optimize the structure for small target objects. Third, we have designed an efficient feature extraction module (EFEM) that can effectively enhance the feature extraction capability of the backbone network for complex environmental information. Finally, to reduce the computational overhead of the model, we have designed a feature fusion lightweight strategy (FFLS) in the neck part, which significantly reduces the computational and parametric quantities of the model. The results of the LMANet on the VisDrone-2021DET and HIT-UAV datasets demonstrate a 4.7% and 2.3% improvement in mean average precision (mAP), respectively, compared to the benchmark model. Additionally, the model's parameters and computation are reduced by 79.3% and 12.6%, respectively.