Abstract: Addressing the interference caused by similar-shaped objects in remote sensing images during road extraction, a residual attention encoder-decoder network (RAED-Net) is proposed. The encoder network of RAED-Net employs an improved channel attention residual module to extract local and global features from the input image. This module adaptively adjusts the weights of channel feature maps, enhancing the focus on important channel information and reducing background interference. In the decoder network, a strip convolution module is introduced to improve cross-channel information interaction during the upsampling process and enhance the ability to recover detailed road edge information, thereby improving the accuracy of road extraction results in complex environments. Comparative experimental results on two different types of public datasets demonstrate that RAED-Net can accurately extract road information, mitigate the interference caused by similar-shaped objects during road extraction, and achieve the best overall results with the smallest number of parameters. Especially on the mini DGRD dataset, which is fully annotated and highly complex, RAED-Net achieves improvements of 3.53%, 5.76%, and 2.21% in F1-score, IoU, and mIoU, respectively, compared to the second-best network.

Key words: remote sensing image, road extraction, encoder-decoder network, channel attention