ConvNeXt embedded U-Net for semantic segmentation in urban scenes of multi-scale targets

Abstract

Semantic segmentation of urban scenes is essential in urban traffic analysis and road condition information acquisition. The semantic segmentation model with good performance is the key to applying high-resolution urban locations. However, the types of these images are diverse, and the spatial relationships are complex. It is greatly affected by weather and light. Objects of different scales pose significant challenges to image segmentation of urban scenes. The existing semantic segmentation is mostly solved from the target scale and superpixel methods. Our research mainly fills the gap in image segmentation field of ConvNeXt fusion U-Net pyramid network model in specific urban scenes. These methods could be more accurate. Therefore, we propose the multi-scale fusion deformation residual pyramid network model method in this paper. This method captures features of different scales and effectively solves the problem of urban scene image segmentation of memory scenes by objects of different scales. We construct a spatial information interaction module to reduce the semantic ambiguity caused by complex spatial relations. By combining spatial and channel characteristics, a series of problems caused by weather and light can be alleviated. We verify the improved semantic segmentation model on the Cityscape dataset. The experimental results show that the method achieves 84.25% MPA and 75.61% MIoU. Our improved algorithm, ConvNeXt embedding in the U-Net algorithm architecture, is named Conv-UNet. The improved method proposed in this paper is superior to other methods in the semantic segmentation of urban scenes. The main advantage of this algorithm is to explore the specific loss function and segmentation strategy suitable for urban scene in the face of the complexity and diversity of urban scene images.