RPLFDet: A Lightweight Small Object Detection Network for UAV Aerial Images With Rational Preservation of Low-Level Features
Ruopu Wang, Chuan Lin, Yongjie Li · IEEE Transactions on Instrumentation and Measurement · 2025
Deep learning-based object detection has achieved great success. However, small object detection remains a challenging task on unmanned aerial vehicles (UAVs) platforms with limited computational resources. Using high-resolution input images or reducing the down-sampling rate of the network can preserve low-level image features, thereby improving the detection performance for small objects. However, these methods will increase the computational cost of the network, resulting in a contradiction between improving the detection performance and reducing the computational cost. To address this dilemma, we propose a rational preservation of low-level features object detection model (RPLFDet). In this network, we introduce a rational stride convolution (RSConv), which allows for non-integer down-sampling rates. RSConv reduces the model’s down-sampling rate while keeping computational cost manageable. To further enhance the efficient processing of low-level features and reduce computational costs, we designed a down-sampling residual (DsR) block. The DsR block reduces spatial redundant information in high-resolution feature maps. Additionally, we propose a novel difference set balanced intersection over union (DSB-IoU) loss to improve the accuracy of small object bounding box regression. Experimental results demonstrate that, considering detection accuracy, computational cost, and model parameter size, our model achieves state-of-the-art performance.