图像自寻的弹药目标检测方法综述

弹载图像目标检测方法是实现图像自寻的弹药“发射后不管”、对目标进行自主打击的关键技术。弹药图像自寻的面临着成像环境恶劣，目标特性变化快，对算法体积、速度要求苛刻等问题。围绕弹载目标检测难点问题进行综述，将基于深度学习的目标检测方法区分为基于候选框、无候选框和基于transformer的方法，回顾了各类方法主要研究进展；对特征提取网络轻量化、预测特征图增强、非极大值抑制后处理算法、训练中样本均衡、模型压缩等弹载图像目标检测模型部署中的关键技术进行了梳理；对比了典型目标检测方法在ImageNet、COCO及弹载图像目标数据集上的性能，并对未来发展进行展望。

Review on Target Detection of Image Homing Ammunition

The onboard image target detection method is the key technology to realize the autonomous attack on the target by the “fire-and-forget” image homing ammunition. At present, the image homing of ammunition is faced with some problems, such as bad imaging environment, rapid change of targets' characteristics, and strict requirements for algorithm volume and speed. Firstly, the target detection methods based on deep learning are divided into methods based on anchor box, methods without anchor box and methods based on transformer, and the main technical progress of various methods is reviewed. Then, the key technologies in onboard image target detection model deployment, such as lightweight feature extraction network, enhancement of feature map for prediction, non-maximum suppression post-processing algorithm, sample equalization in training, and model compression, are studied. Finally, the performances of the typical detection algorithms on ImageNet, COCO and datasets for onboard image are compared, and the possible development in the future is looked into.