基于YOLOv4-tiny的遥感图像飞机目标检测技术研究

Research on remote sensing image aircraft target detection techonlogy based on YOLOv4-tiny

针对传统遥感图像飞机目标检测算法在复杂背景下存在检测准确度和检测召回率较低的问题,基于深度学习中YOLOv4-tiny提出一种遥感图像飞机目标检测算法。根据YOLOv3和YOLOv4的网络结构对YOLOv4-tiny的网络结构进行改进,将原算法中的CSP特征提取网络强化,使其特征提取能力增加;使用Mish激活函数替换原激活函数Leaky ReLU,以获取更好的泛化性;添加了空间金字塔池化模块,缓解网络对目标尺度的敏感程度。实验结果表明:在常规高质量、过度曝光的停机坪、登机口干扰和雾天影响的遥感图像测试中,改进后的算法都有很优秀的检测效果,最终统计检测准确度为98.49%,较原算法提升了1.79%,召回率为97.19%,提升了23.2%,速度达到8.77ms。检测效果有显著提升,能够满足实时性要求。

Aiming at the problem of low detection accuracy and detection recall rate of traditional remote sensing image aircraft target detection algorithm in complex background, a remote sensing image aircraft target detection algorithm based on YOLOv4-tiny in deep learning is proposed. First, according to the network structure of YOLOv3 and YOLOv4, the network structure of YOLOv4-tiny is improved, and the CSP feature extraction network in the original algorithm is strengthened to increase its feature extraction ability. Then, use the Mish activation function to replace the original activation function Leaky ReLU to obtain better generalization. Finally, a spatial pyramid pooling module is added to alleviate the sensitivity of the network to the target scale. The experimental results show that: in the conventional high-quality,over-exposed tarmac,boarding gate interference and foggy remote sensing image test,the improved algorithm has excellent detection results,and the final statistical detection accuracy is 98.49%.Compared with the original algorithm,an increase of 1.79%,a recall rate of 97.19%,an increase of 23.2%,and a speed of 8.77 ms.The detection effect has been significantly improved and can meet realtime requirements.