基于小样本学习的炸点目标检测  

Burst Point Detection Based on Small Sample Learning

作　　者：邹伊 雷志勇[1] ZOU Yi;LEI Zhiyong(School of Electronic and Information Engineering,Xi’an Technological University,Xi’an 710021,China)

机构地区：[1]西安工业大学电子信息工程学院,西安710021

出　　处：《自动化与仪表》2024年第3期93-96,102,共5页Automation & Instrumentation

摘　　要：针对武器测试中炸点目标检测存在的误检、错检问题,提出一种融合自注意力机制、底层信息和解冻权重的两阶段微调的小样本学习方法来进行改进。首先将TFA网络中的FPN替换成带有自注意力机制的AC-FPN网络,并且在金字塔结构部分将底层输出送入顶层,构建一个全新的主干提取网络。然后在对整个网络解冻网络权重,使得新的数据集在整个网络上进行训练。为了验证所提算法,在自制炸点数据集上进行训练和测试,最终该方法的AP为55.2%,比原方法明显提高34.2%,对炸点形状的识别有更好的结果,能更好地满足实际要求。并在Pascal VOC数据集上进行了实验,结果表明该算法的有效性。Aiming at the problems of misunderstanding and wrong inspection in the target test of explosive points in weapon tests,a few-shot learning method with a two-stage fine-tuned samples that integrates self-attention mechanisms,underlying information,and frozen weights to thaw weights was proposed to improve.Firstly,the FPN in the TFA network was replaced with the AC-FPN network with attention mechanism,and the bottom output was fed to the top layer in the pyramid structure to build a new backbone extraction network.The network weights were then unfrozen for the entire network so that the new dataset was trained on the entire network.In order to verify the proposed algorithm,training and testing are carried out on the self-made explosion point data set,and the obtained AP is 55.2%,which is significantly improved by 34.2%compared with the TFA algorithm.It has better results in the recognition of the shape of the explosion point,and can better meet the actual requirements.The experiment is carried out on the Pascal VOC dataset,and the results show the effectiveness of the algorithm.

关 键 词：炸点检测 小样本学习 Faster R-CNN 目标检测 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]