Abstract: Road extraction can help us better understand the urban environment and is an important part of urban transportation and planning.With the development of deep learning and computer vision,the use of deep learning-based semantic segmentation algorithm to extract roads from remote sensing images has become increasingly mature.However,existing deep learning road extraction algorithms suffer from slow extraction speed and susceptibility to background environmental factors,resulting in missed segmentation and discontinuity.To address these issues,a lightweight algorithm called CE-DeepLabv3+ based on ECANet attention mechanism and cascade atrous spatial pyramid pooling module is proposed.Firstly,the main feature extraction network is replaced with the lightweight MobileNetv2 to reduce parameter volume and improve model execution speed.Secondly,the convolution layers of the atrous spatial pyramid pooling module are further expanded to increase the receptive field,and different feature layers are cascaded to enhance semantic information reuse,thereby improving the ability to extract fine-grained features.Thirdly,the ECANet attention mechanism is added to suppress environmental interference and focus on road information.Finally,an improved loss function is used for training to address the impact of road and background sample imbalance on model performance.Experimental results show that the improved algorithm achieves excellent performance,with significant improvements in segmentation efficiency and accuracy compared to the original DeepLabv3+ algorithm.

Key words: Semantic segmentation, Remote sensing images, Road extraction, Attention mechanism, DeepLabv3+, Cascade atrous spatial pyramid pooling