面向无人机影像小目标检测的轻量化算法OA

To address the insufficient feature extraction of existing detection models for tiny targets,this paper proposes a lightweight small object detection model named PRSU-YOLO.First,a Pinwheel-shaped Convolutional Adaptive Module is designed to enhance the directional extraction capability of subtle features.Second,a Reparameterized Spatial-Channel Convolution Module is constructed to optimize multi-scale representation through dynamic feature reconstruction.Third,a Small Object Detection Branch is embedded in the neck network to establish an enhancement pathway for high-resolution detail features.Finally,a Scale-based Dynamic Intersection over Union loss function is introduced,enabling the model to adaptively adjust the bounding-box regression strategy.With only a 14.6 GFLOPS increase in computational complexity,the proposed model achieves an mAP@0.5 of 37.4%on the VisDrone2019,representing a significant improvement of 4.4%over the baseline.On the TinyPerson,it attained a precision of 34.5%,which is an increase of 3.5%compared to the baseline.The experimental results demonstrate that the model significantly enhances detection capability while effectively controlling computational cost,providing an effective solution for small object detection tasks in UAV-based ground observation scenarios.