Machine learning is used to accomplish artificial intelligence tasks such as filling in missing information and image repair. One popular application is repairing images that are obscured, for example, when clouds block aerial views of buildings. While this can be done manually, it’s very time-consuming; even the available machine learning algorithms require many training images before they work, so improving the representation of 3D models using photographs requires additional steps.
In an attempt to improve the accuracy of automatically generated datasets, researchers at Osaka University have applied a machine learning method called generative adversarial networks (GANs). GANs pit two different algorithms against each other. One algorithm (generative network) creates images without clouds, and the second (discriminative network) uses a convolutional neural network to distinguish between digitally repaired pictures and actual images without clouds. As time progresses, both networks improve their ability to create realistic images with clouds digitally removed. The researchers explain that the generative network is trained to fool a discriminative neural network into thinking an image is real.
To automatically generate digital masks that overlay the reconstructed buildings over clouds, the research team used 3D virtual models with photographs from an open-source dataset as input. The model could detect buildings with an intersection over the union of 0.651, which measures how accurately the reconstructed area corresponded to the actual area. This method has many potential applications, such as improving the quality of medical images obscured areas.