针对复杂背景下红外场景对比度低、特征不足、细节不清而导致的目标检测效率低的问题,在 YOLOv5s模型基础上通过创建 TCC(two-way convolution and Concat)模块并引入华为Ghost模块,提出了一种基于改进YOLOv5s模型的红外弱小目标检测方法.首先,结合红外图像的低级语义特征,采取二路卷积和多尺度思想创建了TCC模块,提升了特征提取的全面性;接着,为进一步简化网络结构、减少网络参数量,引入轻量化Ghost模块改进了SPP池化层和CSP2卷积网络;最后,以无人机为实验对象,构建了白天和夜间不同背景条件下的红外弱小目标数据集,实验验证了本文改进算法的有效性.结果表明:改进后的YOLOv5s模型在较少损失帧频的情况下,检测精度提升了 1.34%,平均精度均值(mean average precision,mAP)提升了 2.26%,优于YOLOv4-tiny和YOLOv7-tiny两种轻量化模型,并与YOLOv8s模型精度相当,但模型参数量仅为YOLOv8s模型的53%,完全可以满足嵌入式设备部署的需求.
Infrared dim target detection method based on improved YOLOv5s model
To solve the low accuracy of target detection caused by low contrast,insufficient features,and unclear details in complex infrared scenes,an improved infrared dim target detection method based on YOLOv5s model was proposed by creating a two-way convolution and Concat(TCC)module and introducing the Huawei Ghost module.Firstly,combining the low-level semantic features of infrared images,a TCC module was created using two convolution and multi-scale thought,which improved the comprehensiveness of feature extraction.Then,to simplify the network structure and reduce the number of parameters,a lightweight Ghost module was introduced to improve the SPP pooling layer and CSP2 convolutional network.Finally,using unmanned aerial vehicles(UAVs)as experimental objects,a dataset of infrared dim targets was constructed under various meteorological conditions during day and night,verifying validity of the improved algorithm.The results show that the detection accuracy of the improved YOLOv5s model is increased by 1.34%,and the mean average precision(mAP)is increased by 2.26%,which is superior to YOLOv4-tiny and YOLOv7-tiny models.It has the same accuracy as YOLOv8s model,but the number of model parameters is only 53%of the YOLOv8s model,which fully meets the needs of embedded device deployment.
target detectioninfrared dim targetYOLOv5sTCC moduleGhost module
NETL
NSTL
万方数据
