SECOND-Order Encoder and Restore Detail Decoder Network for Image Semantic Segmentation

Abstract

Traditional convolution and pooling operations in the previous semantic segmentation methods will cause the loss of feature information due to limited receptive field size. They are insufficient to support an accurate image prediction result. To solve this problem, Firstly, we design a Second-Order Encoder to enlarge the feature receptive field and capture more semantic context information. Secondly, we design a Restore Detail Decoder to focus on processing the spatial detail information and refining the object edges. The experiments verify the effectiveness of the proposed approach. The results show that our method achieves competitive performance on two datasets, including PASCAL VOC2012 and Cityscapes with the mIoU of 80.13% and 76.31%, respectively.