SMAPGAN: Generative Adversarial Network-Based Semisupervised Styled Map Tile Generation Method

Traditional online map tiles, which are widely used on the Internet, such as by Google Maps and Baidu Maps, rendered from vector data. The timely updating of tiles data, for generation is time-consuming, a difficult mission. Generating over time remote sensing images relatively simple can be performed quickly without However, this approach to challenging or even impossible. Inspired image-to-image translation (img2img) techniques based generative adversarial networks (GANs), we proposed semisupervised styled GANs (SMAPGAN) model generate directly images. In model, designed learning strategy pretrain SMAPGAN rich unpaired samples fine-tune it limited paired in reality. We also image gradient L1 loss structure tile with global topological relationships detailed edge curves objects, important cartography. Moreover, structural similarity index (ESSI) metric evaluate quality consistency between generated ground truth. experimental results show that outperforms state-of-the-art (SOTA) works according mean squared error, index, ESSI. Also, gained higher approval than SOTA human perceptual test visual realism Our work shows new tool excellent potential producing tiles. implementation available at https://github.com/imcsq/SMAPGAN.