Semantic segmentation and colorization of grayscale aerial imagery with W‐Net models

Abstract

AbstractThe semantic segmentation of remotely sensed aerial imagery is nowadays an extensively explored task, concerned with determining, for each pixel in an input image, the most likely class label from a finite set of possible labels. Most previous work in the area has addressed the analysis of high‐resolution modern images, although the semantic segmentation of historical grayscale aerial photos can also have important applications. Examples include supporting the development of historical road maps, or the development of dasymetric disaggregation approaches leveraging historical building footprints. Following recent work in the area related to the use of fully‐convolutional neural networks for semantic segmentation, and specifically envisioning the segmentation of grayscale aerial imagery, we evaluated the performance of an adapted version of the W‐Net architecture, which has achieved very good results on other types of image segmentation tasks. Our W‐Net model is trained to simultaneously segment images and reconstruct, or predict, the colour of the input images from intermediate representations. Through experiments with distinct data sets frequently used in previous studies, we show that the proposed W‐Net architecture is quite effective in colouring and segmenting the input images. The proposed approach outperforms a baseline corresponding to the U‐Net model for the segmentation of both coloured and grayscale imagery, and it also outperforms some of the other recently proposed approaches when considering coloured imagery.