Lightweight Semantic Segmentation Network Based on Attention Coding

To address the issues of high computational complexity and large memory footprint of the attention map of the self-attention mechanism and to improve the performance of the semantic segmentation network, we propose a lightweight network based on attention coding. The network uses an adaptive positional attention module and global attention upsampling module to encode and decode long-range dependency information, respectively. When calculating the attention map, adaptive positional attention module excludes useless basis sets and context information is obtained. A global attention upsampling module uses global context information to guide low-level features to reconstruct high-resolution images. Experimental results show that the segmentation accuracy of the network on the PASCAL VOC2012 verification set reaches a value of 84. 99o. Compared with dual attention network, which has a similar segmentation accuracy, the giga floating-point operations per second and the GPU memory of the network are reduced by 16.9% and 12.9%, respectively.