Research on lightweight real-time semantic segmentation based on attention mechanism

Abstract

Modern approaches to real-time semantic segmentation algorithms often sacrifice spatial resolution for real-time inference speed, which results in poor performance. Based on this, in this paper, we propose a lightweight real-time semantic segmentation network based on global attention mechanism by improving STDCNet. First, combined with asymmetric convolution, the short-term dense stitching module is light-weighted, and the global correlation of features is enhanced through global attention. Second, the edge branch can effectively filter out the boundary-independent information in the semantic features through the edge feature fusion module, and restore the lost detail information in the decoding stage. Finally, the improved loss function ensures that the network can update parameters in a direction that is conducive to small object segmentation. Tests and analysis on the Cityscapes dataset show that the lightweight contextual attention mechanism achieves 74.9% mIoU and 118.6FPS real-time semantic segmentation inference speed.