针对遥感目标的多尺度、多样性、背景复杂等特点，为提升YOLOv4算法的检测速度和平均精度，提出了一种基于YOLOv4模型改进的遥感目标检测算法，首先，用Mobilenetv2替换YOLOv4的主干特征提取网络，减少参数数量，提升检测速度;其次，在Mobilenetv2的残差网络中嵌入新型注意力机制CoordAttention模块，捕获方向感和位置感知的信息，精准定位和识别感兴趣的目标;最后，借鉴Inception的思想在颈部特征增强网络处添加改进过的RFB模块，增强感受野并且提升网络的特征融合能力。研究表明，论文提出的MCR-YOLOv4(Mobilenetv2-CoordAttention-RFB-You Only Look Once)算法相比于原YOLOv4算法，模型大小减少了45。27 M，平均精度提高了1。03%，检测速度提高了56帧/s，更适用于对复杂遥感目标的检测。
A Remote Sensing Target Detection Algorithm Based on Improved YOLOV4
Aiming at the multi-scale,diversity and complex background of remote sensing targets,in order to improve the de-tection speed and average accuracy of YOLOv4 algorithm,a new algorithm for remote sensing object detection based on YOLOv4 model is proposed.Firstly,the backbone feature extraction network of YOLOv4 is replaced with Mobilenetv2 to reduce the number of parameters and improve the detection speed.Secondly,novel attention mechanism CoordAttention modules embedded in Mobile-netv2's residual network capture information on sense of orientation and position perception,accurately locate and identify targets of interest.Finally,based on the idea of Inception,an improved RFB module is added to the neck feature enhancement network to en-hance the receptive field and enhance the feature fusion capability of the network.It is shown that the proposed MCR-YOLOv4(Mo-bilenetv2-CoordAttention-RFB-You Only Look Once)algorithm reduces the model size in 45.27 M,1.03%improvement in aver-age accuracy and 56 frames/s compared to the original YOLOv4 algorithm,and is more suitable for the detection of complex remote sensing targets.
digital image processingMobilenetv2remote sensing targettarget detectionYOLOv4
国家科技期刊平台
NSTL
万方数据
