摘要
Abstract
In order to improve the target detection accuracy of UAV images,an improved UAV image detection algorithm named DEH-YOLOv10 was proposed to address the problems of missed detections and false detections caused by large variations in target scales,small target sizes,and background noise interference in UAV images.The algorithm improved the C2f module by introducing a dilated residual network(DWR)to enhance the Bottleneck structure and replaced the second convolution module to expand the receptive field and strengthen the network's ability for multi-scale feature extraction.An efficient multi-scale attention(EMA)module was incorporated into the neck network,which improved the model's adaptability to diverse inputs through channel reshaping,grouping,and cross-space learning,while reducing the interference of noise in the detection process.A new shallow detection head,Head4,was added to mitigate the negative effects caused by multi-layer downsampling in the network,thereby improving small target detection accuracy.Comparison experiments and ablation studies on the VisDrone2019 dataset showed that DEH-YOLOv10 improved mAP0.5 by 5.6%and mAP0.5-0.95 by 3.8%compared to the baseline network.These improvements reduced missed and false detections in UAV target detection to some extent and significantly enhanced the overall accuracy of UAV target detection.关键词
无人机图像/目标检测/多尺度特征提取/小目标检测/扩张式残差/高效多尺度注意力Key words
unmanned aerial imagery/object detection/multi-scale feature extraction/small object detection/dilated residuals/efficient multiscale attention分类
信息技术与安全科学
