基于聚类式区域生成的全卷积目标检测网络

Full-Convolution Object Detection Network Based on Clustering Region Generation

基于区域的全卷积网络(R-FCN)的区域生成网络(RPN)沿用了更快速区域卷积神经网络(Faster R-CNN)的RPN。针对RPN先验框的大小与数量均需人为固定,生成的建议区域过多等问题,将聚类思想应用到RPN中,改进先验框的生成方式,提出了基于聚类式区域生成的全卷积目标检测网络。通过对训练样本的真实框进行KMeans聚类得到先验框的最适大小和最佳数量,取代原本人为固定选取先验框的方式。此外,为增强模型的泛化能力,在改进后的R-FCN上使用ResNet基础网络,采用困难样本挖掘方法进行训练。实验结果表明,相较于R-FCN等方法,该聚类区域全卷积目标检测网络得到的检测结果在精度和速度上都得到了明显的提升。

Region proposal networks(RPN)in region-based full-convolutional networks(R-FCN)follow the RPN of faster region convolutional neural networks.In this paper,a full-convolution object detection network based on clustering region generation is proposed to solve the problems of the artificially fixed sizes and quantities of anchor boxes and excessively generated proposals.K-means clustering on the ground-truth box of the training samples is used to optimize the sizes and numbers of the anchor boxes in order to replace the fixed boxes in the R-FCN.Furthermore,to enhance the generalization ability of the model,an online hard example mining is used to train the datasets based on the backbone network of ResNet.The experimental results show that the accuracy of the detection results of the proposed algorithm is significantly higher than that of the R-FCN.