Semantic Relocation Parallel Network for Semantic Segmentation

Semantic segmentation is an essential issue in the computer vision field, the difficulty of which lies in the accurate prediction of the pixel level and the edge division of similar objects. The encoder-encoder structure is widely used in many methods to capture the global information of semantic objects. However, continuous subsampling causes irreversible loss of spatial information of the feature map. A parallel semantic relocation (SRPNet) based network is proposed. Specifically, a high-resolution global spatial path is designed to extract rich spatial information in which feature maps have high resolution. In feature extraction path, a powerful feature extractor is used to expand the receptive field by fast subsampling. Besides, a semantic relocation module (SRM) is designed to compensate for the lack of context information caused by multiple subsamples. Dice loss is employed to alleviate the imbalance of positive and negative samples in the dataset and obtain better segmentation performance. Finally, the proposed network is evaluated on the Cityscapes and CamVid dataset. The results show that SRPNet can improve the previous best result by approximately 3.1% and 1.8% measured by mIoU on the CamVid dataset and the Cityscapes dataset, respectively. © 2022, Beijing China Science Journal Publishing Co. Ltd. All right reserved.