Machine learning techniques for high dimensional data

This thesis presents data processing techniques for three different but related application areas: embedding learning for classification, fusion of low bit depth images and 3D reconstruction from 2D images. For embedding learning for classification, a novel manifold embedding method is proposed for the automated processing of large, varied data sets. The method is based on binary classification, where the embeddings are constructed so as to determine one or more unique features for each class individually from a given dataset. The proposed method is applied to examples of multiclass classification that are relevant for large scale data processing for surveillance (e.g. face recognition), where the aim is to augment decision making by reducing extremely large sets of data to a manageable level before displaying the selected subset of data to a human operator. In addition, an indicator for a weighted pairwise constraint is proposed to balance the contributions from different classes to the final optimisation, in order to better control the relative positions between the important data samples from either the same class (intraclass) or different classes (interclass). The effectiveness of the proposed method is evaluated through comparison with seven existing techniques for embedding learning, using four established databases of faces, consisting of various poses, lighting conditions and facial expressions, as well as two standard text datasets. The proposed method performs better than these existing techniques, especially for cases with small sets of training data samples. For fusion of low bit depth images, using low bit depth images instead of full images offers a number of advantages for aerial imaging with UAVs, where there is a limited transmission rate/bandwidth. For example, reducing the need for data transmission, removing superfluous details, and reducing computational loading of on-board platforms (especially for small or micro-scale UAVs). The main drawback of using low bit depth imagery is discarding image details of the scene. Fortunately, this can be reconstructed by fusing a sequence of related low bit depth images, which have been properly aligned. To reduce computational complexity and obtain a less distorted result, a similarity transformation is used to approximate the geometric alignment between two images of the same scene. The transformation is estimated using a phase correlation technique. It is shown that that the phase correlation method is capable of registering low bit depth images, without any modi�cation, or any pre and/or post-processing. For 3D reconstruction from 2D images, a method is proposed to deal with the dense reconstruction after a sparse reconstruction (i.e. a sparse 3D point cloud) has been created employing the structure from motion technique. Instead of generating a dense 3D point cloud, this proposed method forms a triangle by three points in the sparse point cloud, and then maps the corresponding components in the 2D images back to the point cloud. Compared to the existing methods that use a similar approach, this method reduces the computational cost. Instated of utilising every triangle in the 3D space to do the mapping from 2D to 3D, it uses a large triangle to replace a number of small triangles for flat and almost flat areas. Compared to the reconstruction result obtained by existing techniques that aim to generate a dense point cloud, the proposed method can achieve a better result while the computational cost is comparable