Ghost-UNet: An Asymmetric Encoder-Decoder Architecture for Semantic Segmentation from Scratch

Abstract

One of the most important key points in the intelligent transportation systems is scene understanding of the known and unknown surrounding environment to achieve a safe driving for smart mobile robots and cars. Semantic segmentation can address most of the perception needs of mobile robots and Intelligent Vehicles (IV). There are several deep learning approaches based on Convolutional Neural Network (CNN) for semantic segmentation. Most of these techniques have been designed on a pretrained network base and loading a specific weight file is necessary for them. In this paper, we propose a deep architecture for semantic segmentation from scratch based on an asymmetry encoder- decoder architecture using Ghost-Net and U-Net which we have called it Ghost-UNet. This model can be used for precise segmentation using a combination of low-level spatial information and high-level feature maps. We focus our work on outdoor datasets to evaluate the proposed model which is tested on the Cityscapes dataset. The proposed model has good pixel accuracy and mean Intersection over Union (mIoU) compared with other valid literature.