Localising and Reconstructing Drill Holes in 3D Objects using Machine Learning

Sammanfattning: To move towards an increasingly automated future, machines are becoming smarter and can now solve tasks only humans could before. In this thesis project, the possibility to find, classify and reconstruct holes in three dimensional objects using machine learning is explored. To achieve this, three dimensional convolutional neural networks are used as a method for semantic segmentation. To combat limited GPU memory and training time, a region-based network was created, this network used smaller regions of the 3D objects to process the image in parts, and thereby evade the memory barrier of the GPU, create reconstructions with a higher resolution, and lower training time. The results show that 3D semantic segmentation is possible and is a promising method for reconstruction of features in 3D objects. However, the thesis work also highlights the importance of a qualitative dataset that is a good representation of the data that is intended to be used with the models.