A lightweight attention-guided semantic segmentation network using a group convolution pyramid

Abstract

Semantic segmentation is a key part of computer vision tasks, but its application is greatly limited by memory capacity and computational cost. This paper proposes a lightweight attention-guided semantic segmentation network (LAGNet) that adopts a joint group convolution pyramid strategy. Specifically, we introduced a lightweight symmetric double attention module (LS-DAM) and an adaptive selection interaction module (ASIM). In the LS-DAM, the self-attention adopts adaptive maximum pooling and average pooling instead of its previous form, in which the pixel-by-pixel similarity is calculated. Moreover, The ASIM has the capability to seamlessly integrate both high-level semantic information and low-level geometric information, leading to a remarkable enhancement in the performance of semantic segmentation. We evaluate our proposed LAGNet model using mean intersection over union (mIoU) on PASCAL VOC 2012 and Cityscapes, two commonly used datasets in the field of semantic segmentation, and finally achieve state-of-the-art performance of 81.73% and 81.90%, respectively.