Information embedding and retrieval in 3D printed objects

Deep learning and convolutional neural networks have become the main tools of computer vision. These techniques are good at using supervised learning to learn complex representations from data. In particular, under limited settings, the image recognition model now performs better than the human baseline. However, computer vision science aims to build machines that can see. It requires the model to be able to extract more valuable information from images and videos than recognition. Generally, it is much more challenging to apply these deep learning models from recognition to other problems in computer vision. This thesis presents end-to-end deep learning architectures for a new computer vision field: watermark retrieval from 3D printed objects. As it is a new area, there is no state-of-the-art on many challenging benchmarks. Hence, we first define the problems and introduce the traditional approach, Local Binary Pattern method, to set our baseline for further study. Our neural networks seem useful but straightfor- ward, which outperform traditional approaches. What is more, these networks have good generalization. However, because our research field is new, the problems we face are not only various unpredictable parameters but also limited and low-quality training data. To address this, we make two observations: (i) we do not need to learn everything from scratch, we know a lot about the image segmentation area, and (ii) we cannot know everything from data, our models should be aware what key features they should learn. This thesis explores these ideas and even explore more. We show how to use end-to-end deep learning models to learn to retrieve watermark bumps and tackle covariates from a few training images data. Secondly, we introduce ideas from synthetic image data and domain randomization to augment training data and understand various covariates that may affect retrieve real-world 3D watermark bumps. We also show how the illumination in synthetic images data to effect and even improve retrieval accuracy for real-world recognization applications.