摘要
深度学习技术在目标检测领域取得了显著的成果,但是相关模型在样本量不足的条件下难以发挥作用,借助小样本学习技术可以解决这一问题。本文提出一种新的小样本目标检测模型。首先,设计了一种特征学习器,由Swin Transformer模块和PANET模块组成,从查询集中提取包含全局信息的多尺度元特征,以检测新的类对象。其次,设计了一种权重调整模块,将支持集转换为一个具有类属性的权重系数,为检测新的类对象调整元特征分布。最后在ImageNet-LOC、PASCAL VOC和COCO三种数据集上进行实验分析,结果表明本文提出的模型在平均精度、平均召回率指标上相对于现有的先进模型都有了显著的提高。
Deep learning technology has achieved remarkable results in the field of target detection,but related models are difficult to function under the condition of insufficient sample size.With the help of few-shot learning technology,a new few-shot object detection model is proposed.First,a feature learner is designed,consisting of a Swin Transformer module and a PANET module,to extract multi-scale meta-features containing global information from the query set to detect new class objects.Second,a weight adjustment module is designed to convert the support set into a weight coefficient with class attributes to adjust the meta-feature distribution for detecting new class objects.Finally,experimental analysis is carried out on ImageNet-LOC,PASCAL VOC and COCO datasets.The results show that the model proposed in this paper has a significant improvement in mAP and AR indicators compared to the existing advanced models.
作者
侯玥
王开宇
金顺福
HOU Yue;WANG Kaiyu;JIN Shunfu(School of Information and Engineering,Yanshan University,Qinhuangdao,Hebei 066004,China)
出处
《燕山大学学报》
CAS
北大核心
2023年第1期64-72,共9页
Journal of Yanshan University
基金
国家自然科学基金资助项目(61872311)
秦皇岛市重点研发计划资助项目(202101A015)。
