基于背景抑制和分类校正的小样本目标检测方法
Few shot target detection method based on background suppression and classification correction
蔡伟 1王鑫 1蒋昕昊 1杨志勇 1陈栋1
作者信息
- 1. 火箭军工程大学 导弹工程学院,陕西 西安 710025
- 折叠
摘要
为进一步提升小样本条件下对空中来袭目标的检测识别成功率,提出一种基于背景抑制和分类校正的小样本目标检测方法.首先,针对空中来袭目标背景前景易混淆问题,在区域候选网络前端引入背景抑制模块,通过抑制背景特征和增强前景特征来减轻目标背景对检测的影响;其次,在背景抑制模块后插入特征聚合模块,聚焦目标特征,通过缓解小样本条件导致目标特征难提取、不明显的问题,校正网络模型的分类参数;最后,在检测头网络中引入对比分支,增强了类别内相似性和类别间独特性,缓解来袭目标"类间相似性高,类内差异性大"的问题,实现了对网络分类的进一步校正.实验结果表明,所提出的算法在 1、2、3、5、10 shot实验中均表现最佳,平均精度分别达到28.3%、32.8%、39.9%、42.9%和56.2%,提升了小样本空中来袭目标的检测性能.
Abstract
To further improve the success rate of detecting and identifying airborne targets under few shot conditions,a few shot target detection method based on background suppression and classification correction was proposed.Firstly,aiming at the problem that the background foreground of incoming air targets was easy to confuse,a background suppression module was introduced in the front end of the regional candidate network,which enhanced the foreground features by suppressing the background features and reduced the influence of the target background on detection.Secondly,the feature aggregation module was inserted after the background suppression module to focus on the target features,and to alleviate the problem that the target features were difficult to extract and not obvious due to few shot conditions,so as to correct the classification parameters of the network model.Finally,a contrast branch was introduced into the detection head network for enhancing the similarity within classes and uniqueness between classes,which alleviated the problem of high similarity between classes and large differences within classes of incoming targets,and realized the further correction of the network classification.The experimental results show that the proposed algorithm performs best in the 1,2,3,5 and 10 shot experiments,with average accuracy reaching 28.3%,32.8%,39.9%,42.9%and 56.2%,respectively,which improves the detection performance of few shots airborne incoming targets.
关键词
小样本目标检测/空中来袭目标/背景抑制/分类校正/深度学习Key words
few shot target detection/airborne incoming targets/background suppression/classification correction/deep learning引用本文复制引用
基金项目
173技术领域基金(2021-JCJQ-JJ-0871)
出版年
2024
NETL
NSTL
万方数据