08.06.23 - Bringing History to Life: Advancements in Colorizing Historical Aerial Images
Black and white images are a barrier for a large proportion of the population. Coloring them would make them easier to interpret for the general public, as well as for automatic applications. Examples of applications could be resolving neighborhood conflicts or identifing changes in land cover.
As part of the TIME (hisTorical aerIal iMagEs) project, Farella et. All* have been working on evaluating the recent aerial image colorization solutions, developing a new Hyper-U-NET neural network architecture - a combination of a U-NET-Like network and HyperConnections - and preparing a new benchmark dataset for colorizing historical aerial images. They trained their model on a set of 10’000 aerial images of different scales and environments and showed an improvement over previous methods. Four evaluation metrics from the literature were used to assess the results and compare the models.
Some further conclusions were that the quality of predictions depends on the quality of the input data. Image enhancement and restoration can be performed prior to colorization to aid prediction. Other color spaces can also be tried out to improve the performance of the method.
As the solution developed is open source (https://lnkd.in/euV-R68V), the Swiss Territorial Data Lab (STDL) wanted to test this new deep learning architecture on Swiss aerial images. One can visually appreciate the prediction on some SWISSIMAGE and SWISSIMAGE HIST 1946 samples. For the SWISSIMAGE sample, the original RGB patch is first converted into a black-and-white patch before being used for the prediction.
#historicalimages #colorization #deeplearning #neuralnetwork #data #project
*Farella, Elisa Mariarosaria, Salim Malek, and Fabio Remondino. “Colorizing the Past: Deep Learning for the Automatic Colorization of Historical Aerial Images.” Journal of Imaging 8, no. 10 (October 2022): 269. https://lnkd.in/eWXSCN89