In the master's thesis, we address the problem of shadow removal from orthophotography images, which presents a challenge due to the lack of suitable training data. To tackle this issue, we developed an innovative approach. This involves creating our own dataset by generating data using the Unity environment, where we use the orthophoto as a terrain layer and 3D models of buildings and vegetation to render realistic shadows, followed by training a U-Net neural network. We evaluated the model on various training datasets and neural networks. We found that data quality is crucial for successful learning, as evidenced by our dataset's competitive results. The main contributions of the thesis are a comprehensive training dataset and a trained shadow removal model.
|
