AI-Assisted 3D Human Head Keypoint Matching for Image Pair Alignment

This thesis work is about identifying, through a Computer Vision approach, specific keypoints within MR (”Magnetic Resonance”) volumes of human heads. Such keypoints can be exploited to align the volumes with other targets, such as the Augmented Reality representation of the users' heads or other scanned images. The first phase of the study focused on the identification of training and test datasets, the selection of four non-coplanar points of interest for the alignment, and the identification of the most appropriate neural network architecture to individuate such points. In particular, different 3D CNN ("Convolutional Neural Network") architectures were tested, with different sets of hyperparameters and exploiting multiple volume sampling frequencies. Moreover, we designed a customized data augmentation procedure in order to enlarge the available training samples and better the network generalization capabilities. Once obtained a model with sufficient performances for the alignment, the following steps were the implementation of the prediction phase in JavaScript and, finally, the image-to-image alignment. The dataset collection and keypoints annotation, together with the final model, are an evident contribution of this thesis work to the state of the art