摘要
针对无人机航拍图像检测在复杂场景下难以检测出目标以及对于小目标检测精度低的问题,提出一种适用于无人机航拍图像的目标检测算法.为提高小目标检测精度,引入小目标检测层和基于空洞卷积与注意力机制的特征增强模块,从多尺度特征图中提取更多特征,增强模型对小目标的识别能力.为应对复杂场景,引入前景感知模块,关联与前景相关的上下文信息,减少背景干扰.在VisDrone2021 数据集上实验,与YOLOv5s相比,改进后模型mAP50 提高 5.3%,mAP50:95 提高3.4%.在 960×960 分辨率下训练的模型mAP50 达 47.9%,比YOLOv4 高 4.9%,适用于无人机航拍图像检测.
Abstract
To enhance the model's recognition capability,this paper introduces a small target detection layer and a fea-ture enhancement module based on void convolution and attention mechanism to extract more features from multi-scale feature maps.Additionally,this paper incorporates a foreground awareness module to associate contextual information relat-ed to the foreground and reduce background interference.Experimental results on the VisDrone2021 dataset demonstrate significant improvements over YOLOv5s,with an increase of 5.3%in mAP50 and 3.4%in mAP50:95.Moreover,our model trained at 960×960 resolution achieves an mAP50 of 47.9%,which is 4.9%higher than YOLOv4,making it suitable for UAV aerial image detection.
维普
万方数据