Research on Semantic Segmentation of Nighttime Building Facades Based on Multi-Exposure Data
Abstract：Accurate semantic segmentation of nighttime building facades is crucial for urban nighttime research. The complex and changing lighting conditions in the night lead to difficulties in capturing detailed features of building facades, and traditional segmentation methods often suffer from issues such as blurred boundaries and feature loss. Based on the information complementarity of multi-exposure data, this paper proposes an improved DeeplabV3 algorithm. This algorithm constructs a three-branch backbone network to simultaneously extract features from images with different exposure levels, significantly enhancing the ability to perceive details of nighttime building facades. Experimental results on a drone-based nighttime building facade semantic segmentation dataset show that the introduction of multi-exposure data effectively improves segmentation accuracy. The improved algorithm achieves a 1.0% increase in mean intersection over union (mIoU) compared to DeeplabV3, reaching 92.8%, achieving more accurate boundary segmentation of building facades.
Keywords：Semantic SegmentationDrone Nighttime ImageryMulti-ExposureBuilding Facade
Publication Date：2025-07-31
Pages：4( 6-9 )