MFD-UNet: a complex railway track boundary segmentation algorithm for drone vision based on multi-scale fusion and deformable pyramid blocks

Abstract Segmenting railway tracks from aerial imagery is critical in producing electronic railway maps. To address the challenges of railway track extraction from complex aerial imagery, this paper proposes an algorithm for complex railway track boundary segmentation Algorithm from complex aerial imagery based on MFD-UNet (Multi-scale attention Fusion and Deformable pyramid UNet). Firstly, we proposed a parallel deformable convolutional pyramid residual downsampling module. This module integrates deformable convolution with fixed convolution, dynamically adjusting the receptive field by learning offset values, and enhances the extraction of information at different scales by refining the residual blocks based on a pyramid structure. Secondly, a multi-scale attention fusion module is proposed in the skip connection part to integrate adjacent feature layers and reduce semantic differences. Additionally, a multi-layer optimized cross-entropy loss function is integrated into the decoder to enhance the model’s perception of information at various scales. Experimental results show that our algorithm can effectively segment railway tracks on both our self-built railway dataset and the Deep Globe dataset. Compared to the benchmark model, it has improved by 3.19% and 4.51% respectively, and the accuracy has also increased by 2.38% and 2.95%, achieving a good visual effect.