摘要
一般的目标检测算法均是建立在训练集与测试集相同分布的情况下,但是在应用过程中,经常会出现模型训练的场景与实际使用的场景存在偏差的问题。这样的分布不匹配会导致模型性能的大幅下降。针对新的应用场景,重头训练又需要大量标注数据,这将耗费大量人力与时间,代价昂贵。针对目标检测任务在源域与目标域的分布差异导致的模型性能下降问题,结合域适应的思想,从特征图和检测区域两个层面对目标检测算法Faster R-CNN进行改进,提升目标检测模型在多场景下的检测精度。此外,该模型进一步采用双层ROI-Pooling的方法提升域适应效果。模型采用无监督学习的方法,对两个不同角度下的街道场景数据集进行实验,提升了目标域车辆检测的精度。
General object detection algorithms are based on the same distribution of training set and test set.However,in the application,there is often the problem of deviation between the training scene and the actual scene of the model.Such a mismatch in distribution will lead to a significant drop in model performance.For new application scenarios,re-training model requires a lot of labeling data,which is expensive and difficult.In order to relieve the performance decline of source training model in target domain,we combine the idea of domain adaptation to improve the object detection model Faster R-CNN from two aspects of feature map and detection region,so as to improve the detection accuracy of the object detection model in multiple scenes.In addition,the two-layer ROI-Pooling method is further adopted in this model to enhance the domain adaptation effect.The experiment is made in two different street scene datasets with unsupervised leaning method to improves the accuracy of vehicle detection in the object area.
作者
王翎
孙涵
WANG Ling;SUN Han(School of Computer Science and Technology,Nanjing University of Aeronautics and Astronautics,Nanjing 211106,China)
出处
《计算机技术与发展》
2019年第12期158-161,166,共5页
Computer Technology and Development
基金
中央高校基本科研业务费专项资金(NS2016091)
南京航空航天大学研究生创新基地(实验室)开放基金(kfjj20171602)
