EAR-Net: Efficient Atrous Residual Network for Semantic Segmentation of Street Scenes Based on Deep Learning

Abstract

Segmentation of street scenes is a key technology in the field of autonomous vehicles. However, conventional segmentation methods achieve low accuracy because of the complexity of street landscapes. Therefore, we propose an efficient atrous residual network (EAR-Net) to improve accuracy while maintaining computation costs. First, we performed feature extraction and restoration, utilizing depthwise separable convolution (DSConv) and interpolation. Compared with conventional methods, DSConv and interpolation significantly reduce computation costs while minimizing performance degradation. Second, we utilized residual learning and atrous spatial pyramid pooling (ASPP) to achieve high accuracy. Residual learning increases the ability to extract context information by preventing the problem of feature and gradient losses. In addition, ASPP extracts additional context information while maintaining the resolution of the feature map. Finally, to alleviate the class imbalance between the image background and objects and to improve learning efficiency, we utilized focal loss. We evaluated EAR-Net on the Cityscapes dataset, which is commonly used for street scene segmentation studies. Experimental results showed that the EAR-Net had better segmentation results and similar computation costs as the conventional methods. We also conducted an ablation study to analyze the contributions of the ASPP and DSConv in the EAR-Net.