The thesis aims to train a model that would automatically detect clouds on aerial photographs using deep learning. The main reason for the task was the project of cyclic laser scanning of Slovenia, which includes capturing photographs to create an orthophoto. Because of this it is necessary to process large amounts of images. With any automatization, one avoids manual inspection of images, which is needed to create an orthophoto. In the thesis we tried different architectures. Ultimately, we decided on Faster R-CNN and performed the learning within the open-source Detectron2 library. The model was pre-trained on images from Microsoft's COCO collection from 2017. Due to the variability of the terrain the model needed to be generalized. We achieved this by using data augmentation. We used the techniques of flipping, rotation, contrast change and brightness change. To obtain the best possible model, we periodically evaluated it on the validation dataset during learning and saved intermediate iterations. Finally, we ran it on the test set, which contains 90 images with and without clouds.
|
