A Lightweight High-Resolution RS Image Road Extraction Method Combining Multi-Scale and Attention Mechanism

• Published in: IEEE access Vol. 11; pp. 108956 - 108966
• Main Authors: Wang, Rui, Cai, Mingxiang, Xia, Zixuan
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2023; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Algorithms; ASPP; CAM; Convolutional neural networks; Data dissemination; Data mining; Datasets; Deep learning; Discontinuity; Feature extraction; High resolution; Image edge detection; Image enhancement; Image resolution; Image segmentation; Information sharing; Lightweight; Machine learning; Modules; Neural networks; Remote sensing; Road extraction; Roads; Roads & highways; SAM; Training; United States > US; Massachusetts
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2023.3313390

Road information plays an indispensable role in human society's development. However, owing to the diversity and complexity of roads, it is difficult to obtain satisfactory road-extraction result. Some typical factors, such as discontinuity, loss of edge details, and long-time consumption, have negative impacts on obtaining accurate road information. These problems are particularly prominent during road extraction when high-resolution remote-sensing images are used. To obtain accurate road information, a novel lightweight deep learning neural network was pro-posed in this study by integrating a multiscale module and attention mechanisms. As an excellent multiscale segmentation module, the atrous spatial pyramid pooling was selected to enhance the road extraction ability of remote sensing images. In addition, an attention mechanism was employed to solve the problems of discontinuity and loss of edge details in road extraction, and MobileNet V2 was selected as the backbone of DeepLab V3+ because of its lightweight structure, which can help solve the problem of excessive training time consumption. The experimental verification was carried out on the Ottawa road dataset and the Massachusetts road dataset. Experimental results show that compared with U-Net, SegNet and MDeeplab v3+ networks, the proposed algorithm is the best in IoU, Recall, OA and Kappa. Among them, on the Ottawa road dataset, the OA and Kappa of the algorithm in this paper are 98.92 % and 95.02 %, respectively. On the Massachusetts road dataset, OA and Kappa 98.29% and 89.87%. In addition, the training time was significantly shorter than that of the other deep learning networks. The proposed method exhibited a good performance in road extraction.