艾拓,梁亚玲,杜明辉.基于难负样本挖掘的改进Faster RCNN训练方法[J].计算机科学,2018,45(5):250-254
基于难负样本挖掘的改进Faster RCNN训练方法
Improved Faster RCNN Training Method Based on Hard Negative Mining
投稿时间:2017-03-31  修订日期:2017-06-05
DOI:10.11896/j.issn.1002-137X.2018.05.043
中文关键词:  甚高速区域卷积网络,目标检测,难负样本挖掘,自助采样
英文关键词:Faster RCNN,Object detection,Hard negative mining,Bootstrap sampling
基金项目:本文受广州市科技计划项目(201707010070)资助
作者单位E-mail
艾拓 华南理工大学电子与信息学院 广州510641  
梁亚玲 华南理工大学电子与信息学院 广州510641 ylliang@scut.edu.cn 
杜明辉 华南理工大学电子与信息学院 广州510641  
摘要点击次数: 343
全文下载次数: 246
中文摘要:
      目标检测方法甚高速卷积神经网络(Faster Region-based Convolutional Neural Network,Faster RCNN)在训练过程中存在负样本远多于正样本的问题,即数据集不平衡问题。针对该问题,提出了一个综合定位误差和分类误差的判别函数用于判别难正样本,基于该函数和难负样本挖掘提出了改进的自助采样法,并提出了基于该自助采样的 “五步训练法”用于训练Faster RCNN。与传统的Faster RCNN训练方法相比,五步法加强了对难样本的学习,提高了网络泛化能力,减少了误判;训练出的模型在Pascal VOC 2007数据集上测试的平均正确率均值(mean Average Precision,mAP)提高了2.4%,在FDDB(Face Detection Data Set and Benchmark)相同检出率下误检率降低了3.2%,且边框拟合度更高。
英文摘要:
      In the training process of object detection method named Faster RCNN(Faster Region-based Convolutional Neural Network),there is a data imbalance problem which means that training data contains an overwhelming number of negative examples.Aiming at this problem,a discriminant function was proposed to distinguish hard positive examples,which combines location loss and classification loss.Based on this function and hard negative mining,an improved bootstrap sampling method was proposed.Five-step training method was proposed by introducing the bootstrap sampling into traditional Faster RCNN training.Comparing with the traditional training,this method improves network’s generalization ability,reduces false positive rate,and can learn hard example better.The experimental results show that the model trained by five step attains 2.4% higher mAP(mean Average Precision) on Pascal VOC 2007 dataset,reduces false positive by 3.2% on FDDB(Face Detection Data Set and Benchmark) with the same true positive rate,and gets higher fitting degree of boundary box.
查看全文  查看/发表评论  下载PDF阅读器