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.