C2S-RoadNet: Road Extraction Model with Depth-Wise Separable Convolution and Self-Attention

Abstract

In order to effectively utilize acquired remote sensing imagery and improve the completeness of information extraction, we propose a new road extraction model called C2S-RoadNet. C2S-RoadNet was designed to enhance the feature extraction capability by combining depth-wise separable convolution with lightweight asymmetric self-attention based on encoder and decoder structures. C2S-RoadNet is able to establish long-distance dependencies and fully utilize global information, and it better extracts road information. Based on the lightweight asymmetric self-attention network, a multi-scale adaptive weight module was designed to aggregate information at different scales. The use of adaptive weights can fully harness features at different scales to improve the model’s extraction performance. The strengthening of backbone information plays an important role in the extraction of road main branch information, which can effectively improve the integrity of road information. Compared with existing deep learning algorithms based on encoder–decoder, experimental results on various public road datasets show that the C2S-RoadNet model can produce more complete road extraction, especially when faced with scenarios involving occluded roads or complex lighting conditions. On the Massachusetts road dataset, the PA, F1 score, and IoU reached 98%, 77%, and 72%, respectively. Furthermore, on the DeepGlobe dataset, the PA, F1 score, and IoU reached 98%, 78%, and 64%, respectively. The objective performance evaluation indicators also significantly improved on the LSRV dataset, and the PA, F1 score, and IoU reached 96%, 82%, and 71%, respectively.

Keywords

• remote sensing imagery
• road extraction
• self-attention
• feature fusion