Semantic image segmentation with shared decomposition convolution and boundary reinforcement structure

摘要

Deep convolutional neural networks (DCNNs) have shown excellent performances in the field of computer vision. In this paper, we propose a new semantic image segmentation model, and the two hallmarks of our architecture are the usage of shared decomposition convolution (SDC) operation and boundary reinforcement (BR) structure. SDC operation can extract dense features and increase correlation of features in the same group, which can relieve the grid artifact problem. BR structure combines the spatial information from different layers in DCNNs to enhance the spatial resolution and enrich target boundary position information simultaneously. The simulation results show that the proposed model can achieve 94.6% segmentation accuracy and 76.3% mIOU on PASCAL VOC 2012 database respectively, which verifies the effectiveness of the proposed model.