基于MVT-CTRFBs-YOLOv4的遥感飞机检测研究

Research on Remote Sensing Aircraft Detection Based on MVT-CTRFBs-YOLOv4

针对目前遥感图像检测算法存在实时性差、精度低、召回率低的问题,提出了一种基于移动视觉Transformer(Mobile Vision Transformer,MVT)和语境Transformer感受野模块(Contextual Transformer Receptive Field Blocks,CTRFBs)的改进YOLOv4(You Only Look Once version4)遥感飞机检测算法。首先,为了降低模型参数量,采用轻量级主干网络MVT替换原始CSPDarknet53(Cross Stage Partial Darknet53)主干网络进行特征提取,从而提高检测速度。其次,为了减少小目标丢失问题,引入CTRFBs代替原YOLOv4的空间金字塔池化(Spatial Pyramid Pooling,SPP)结构增大浅层特征层感受野以提升召回率。最后,在原始YOLOv4的路径聚合网络(Path Aggregation Network,PANet)中引入多个SE(Squeeze-and-Excitation)注意力模块,加强小目标特征提取从而提高检测精度。在RSOD和UCAS_AOD数据集上的实验表明,相比其他改进YOLOv4算法,改进模型的检测精度达到94.65%,召回率达到91.55%,模型大小仅为25.95 MB。该算法不仅实现了网络结构的轻量化,而且明显提高了遥感飞机的检测效果。

Aiming at the problems of poor real-time performance,low precision and low recall of current remote sensing image detection algorithms,an improved you only look once version4(YOLOv4)remote sensing aircraft detection algorithm based on mobile vision transformer(MVT)and context transformer receiver field blocks(CTRFBs)is proposed.First of all,in order to reduce the amount of model parameters,the lightweight backbone network MVT is used to replace the original cross stage partial darknet53(CSPDarknet53)backbone network for feature extraction,so as to improve the detection speed.Secondly,in order to reduce the problem of small target loss,CTRFBs are introduced to replace the original spatial pyramid pooling(SPP)structure of YOLOv4 to increase the receptive field of the shallow feature layer to improve the recall rate.Finally,multiple squeeze and extraction(SE)attention modules are introduced into the original YOLOv4 path aggregation network(PANet)to enhance small target feature extraction and improve detection accuracy.Experiments on the data set of RSOD and UCAS_AOD show that compared with other improved YOLOv4 algorithm,the detection accuracy of the improved model reaches 94.65%,the recall rate reaches 91.55%,and the model size is only 25.95 MB.The algorithm not only realizes the lightweight of network structure,but also significantly improves the detection effect of remote sensing aircraft.