RPE-Net: Road Patch Extraction Network for Improving the Integrity of Road Extraction Results from Remote Sensing Images

Deep learning (DP) based road extraction methods often produce fragmented results. Direct optimization of end-to-end DP-based methods requires the design of more complex network structures to enhance the model’s adaptability in complex scenarios. To tackle this challenge, this paper adopts a novel approach that treats the discrepancy (the road breakage part) between road prediction and ground-truth as the extraction target, thereby constructing a highly efficient and lightweight semantic segmentation network, termed the Road Patch Extraction Network (RPE-Net). RPE-Net includes multi-directional striped residual (MDSR) encoder, multi-directional striped pooling (MDSP) units, and multi-directional striped decoder (MDSD). The structure is similar to LinkNet34, but the overall size of the network parameters is just 1.56MB, which is 1/50 of LinkNet34. The post-processing datasets used for training can be semi-automatically generated by the algorithm, and only a small amount of manual intervention can be used for training. A large number of experiments showt hat the post-processing method proposed in this paper has extremely high speed and generalization ability.