面向无人机遥感场景的轻量级小目标检测算法

Lightweight Small Object Detection Algorithm for UAV Remote Sensing Scene

在基于深度学习的目标检测算法中,YOLO算法因兼具速度与精度的优势而备受关注,但是将其应用于无人机遥感领域时存在检测速度较慢、计算资源要求较高、小目标检测精度不佳等问题。为此,提出基于YOLO的轻量级小目标检测算法SS-YOLO。使用轻量的主干网络提升算法的推理速度,根据特征金字塔网络分治思想,加入下采样倍数为4的高分辨特征图P2用于检测微小目标。为解决高分辨率特征图(P2、P3)中语义信息不足的问题,构建结合自适应融合因子的语义增强上采样模块。针对定位损失函数中IoU度量方法对目标尺寸敏感所带来的影响小目标定位精确性的问题,设计结合归一化Wasserstein距离度量方法与中心点距离惩罚项的L_(CNWD)定位回归损失函数。实验结果表明,与YOLOv5s以及最新的YOLOv7-tiny相比,改进后的SS-YOLO模型参数量分别减少了31.3%和20.6%,与YOLOv7-tiny相比,mAP在VisDrone与AI-TOD数据集上分别提升了7.5和7.0个百分点;与YOLOv5s相比,mAP分别提升了2.3和3.6个百分点。当输入图片尺寸为800×800像素时,SS-YOLO的FPS为110帧/s,能够在满足无人机等边缘设备实时检测的同时,显著提升小目标的检测结果。

In the field of deep learning,particularly in object detection algorithms,the YOLO algorithm stands out for its speed and accuracy.However,its application in Unmanned Aerial Vehicle(UAV)remote sensing encounters challenges such as slow detection speed,high computational demands,and decreased accuracy in detecting small objects.To overcome these limitations,this paper introduces SS-YOLO,a lightweight variant of YOLO optimized for small object detection.SS-YOLO utilizes a lightweight backbone network to enhance the algorithm's inference speed.It adopts the divide-and-conquer approach of the Feature Pyramid Network(FPN)and integrates a high-resolution feature map,P2,with a downsampling factor of four,specifically for small target detection.The paper also proposes a semantic enhancement upsampling module combined with adaptive fusion factors to address the semantic information deficiency in high-resolution feature maps(P2,P3).Moreover,SS-YOLO features an innovative L_(CNWD) localization regression loss function.This function merges the Normalized Wasserstein Distance(NWD)measurement method with a center point distance penalty term.This integration effectively addresses the sensitivity of the Intersection over Union(IoU)measurement method to target size in the localization loss function,which impacts the accuracy of small target localization.Experimental results indicate that SS-YOLO surpasses YOLOv5s and YOLOv7-tiny in efficiency.It reduces the parameter count by 31.3%and 20.6%respectively,compared to these models.On the VisDrone and AI-TOD datasets,SS-YOLO shows an increase of 7.5 and 7.0 percentage points in mean Average Precision(mAP),respectively,when compared to YOLOv7-tiny.Against YOLOv5s,the mAP increases by 2.3 and 3.6 percentage points,respectively.Notably,with an input image size of 800×800 pixels,SS-YOLO achieves a Frames Per Second(FPS)of 110 frame/s,demonstrating its capability to significantly improve the detection of small objectts while meeting the real-time detection requirements of edge devices such as drones.