Global-feature Enhanced Network for Fast Semantic Segmentation

摘要

Fast-SCNN is a lightweight, real-time dual-branch semantic segmentation network that can capture both global and local information. Global information is essential for segmentation to enhance the semantic understanding. However, but the current deep network branches of Fast-SCNN have limitations in effectively extracting such information. Finding a balance between speed and precision becomes more important as demand for robotics and self-service systems rises. To solve these issues, we propose a Global-feature Enhanced Network (GFEN) for Fast Semantic Segmentation. In particular, the contextual information of the surrounding pixels on the crossover route is retrieved by adding an attention module to the deep branches of the dual branch network, which efficiently extracts global features and achieves a good compromise between accuracy and speed. The attention module known as Criss-Cross Attention (CCA) is capable of efficiently extracting global features with minimal parameter amounts. Extenstive experiments on The efficacy of our suggested strategy is demonstrated by the Cityscapes dataset, and we have obtained 60.546% mIoU and 7.044G FLOPs, with negligible increase in the number of parameters.