多重金字塔的轻量化遥感车辆小目标检测算法

Small object detection algorithm for lightweight remote sensing vehicles with multiple pyramids

针对遥感车辆检测任务中存在目标尺寸小、背景复杂等问题,提出一种基于多重金字塔和多尺度注意力的轻量级YOLOv5算法。在主干网络中减少下采样次数,提高小目标检测能力,实现轻量化;在颈部中通过重新设计的多重金字塔网络,充分利用不同特征层的信息,增强特征融合能力,并引入改进的多尺度注意力模块,为浅层特征图获得更大的感受野和感兴趣区域;最后使用K-means++聚类算法对目标尺寸进行聚类分析,设计出适合目标的锚框尺度和宽高比。在自建遥感车辆数据集中不仅提升了目标检测精度,而且大大降低参数量。与YOLOv5s相比较,AP_(0.5)%提高了2.3%、AP_(0.5:0.75)%提高了4.3%;参数量降低了65%、模型大小减少了60%。在轻量化的同时有效地提高了小目标的检测精度。

Aiming at the problems of small target size and complex background in remote sensing vehicle detection tasks,a lightweight YOLOv5 algorithm based on multiple pyramids and multi-scale attention is proposed.In the backbone network,the number of downsampling is reduced,the small target detection ability is improved,and light weight is achieved;in the neck,the information of different feature layers is fully utilized through the redesigned multipyramid network to enhance the feature fusion ability.And introduce an improved multi-scale attention module to obtain a larger receptive field and area of interest for the shallow feature map;finally,the K-means++clustering algorithm is used to cluster and analyze the target size,and an anchor frame scale suitable for the target is designed.and aspect ratio.In the self-built remote sensing vehicle dataset,the target detection accuracy is not only improved,but also the parameter quantity is greatly reduced.Compared with YOLOv5s,AP_(0.5)%is increased by 2.3%,AP_(0.5:0.75)%is increased by 4.3%;the number of parameters is reduced by 65%,and the model size is reduced by 60%.It effectively improves the detection accuracy of small targets while reducing weight.