Multi-scale Network and Ghost Attention Head for Semantic Segmentation

Abstract

In view of the existing problems in semantic segmentation, such as insufficient utilization of information at different scales, excessive computing costs for generating redundant information, and ignoring spatial information. In this end, we propose a multi-scale convolution and ghost attention head. The former is inserted into the encoder, which increases the receptive field of the network and makes the network more effective to obtain context information. At the same time, a Ghost Attention Head is designed in the process of decoding, which adopts cheap operation and separable attention to guide the network, so as to solve the problems such as computational cost and ignoring spatial information. Experimental results on CamVid show that the mIoU of the proposed module reaches 71.32%, which is 9.7% higher than that of the traditional semantic segmentation network FCN. Ablation experiments were also carried out to further verify the effectiveness of the added module.