Unsupervised historical map registration by a deformation neural network

Image registration that aligns multi-temporal or multi-source images is vital for tasks like change detection and image fusion. Thanks to the advance and large-scale practice of modern surveying methods, multi-temporal historical maps can be unlocked and combined to trace object changes in the past, potentially supporting research in environmental science, ecology and urban planning, etc. Even when maps are geo-referenced, the contained geographical features can be misaligned due to surveying, painting, map generalization, and production bias. In our work, we adapt an end-to-end unsupervised deformation network that couples rigid and non-rigid transformations to align scanned historical map sheets at different time stamps. To the best of our knowledge, we are the first to use unsupervised deep learning to register map images. We address the sparsity of map features by introducing a loss based on distance fields. When aligning the displaced landmark locations by our proposed method, the results are promising both quantitatively and qualitatively. The generated smooth deformation grid can be applied to vector features directly to align them from the source map sheet to the target map sheet.