Abstract:
A common problem in computer vision is to match corresponding points between images. The success of computer vision has usually relied on having good feature representations, which are usually hand-crafted and thus require huge amounts of prior knowledge and human labor. To the best of my knowledge, feature learning algorithms have not been used for feature matching in the literature. In this thesis, I will present how to use feature learning algorithms to build useful feature representations from aerial images in a biologically-inspired and unsupervised manner. These learned feature representations can then be used to do feature matching tasks. Specifically, I will present two algorithms, Sparse Filtering and convolutional Networks for generating hierarchical representations, in which the information from the lower levels is grouped to establish complex features. These complex and hierarchical representations often lead to performance approaching highly hand-designed computer vision algorithms in feature matching tasks.
