基于有效感受野的区域推荐网络

Region Proposal Network Based on Effective Receptive Field

基于卷积神经网络的目标检测方法通过优化区域推荐达到较高的检测精度.由此,文中提出基于有效感受野的区域推荐网络.在区域推荐网络上引入基于有效感受野的样本匹配方法,强化基于交叠比的样本匹配规则,增强特征信息在区域推荐生成时的表征能力,减少锚定框和区域推荐数目,简化锚定框参数设置.结合快速区域的卷积神经网络检测器后,在Pascal VOC数据集上的检测精度有所提升,这表明文中方法是有效的.

Object detection methods based on convolutional neural network(CNN)optimize region proposal to achieve a higher detection accuracy.Therefore,an effective receptive field(eRF)based region proposal network is proposed.A sample matching method based on eRF is introduced into regional proposal network.Thus,the intersection over union(IoU)based sample matching rule is improved.The representation ability of feature information in the region proposal generation stage is enhanced.The number of region proposal and anchor boxes is greatly reduced.The parameter settings of anchor boxes are also simplified.The detection accuracy on Pascal VOC datasets is improved in combination with Fast R-CNN detector.The effectiveness of proposed method is verified.