Visual Scene Understanding by Deep Fisher Discriminant Learning

Abstract

Modern deep learning has recently revolutionized several fields of classic machine learning and computer vision, such as, scene understanding, natural language processing and machine translation. The substitution of feature hand-crafting with automatic feature learning, provides an excellent opportunity for gaining an in-depth understanding of large-scale data statistics. Deep neural networks generally train models with huge numbers of parameters, facilitating efficient search for optimal and sub-optimal spaces of highly non-convex objective functions. On the other hand, Fisher discriminant analysis has been widely employed to impose class discrepancy, for the sake of segmentation, classification, and recognition tasks. This thesis bridges between contemporary deep learning and classic discriminant analysis, to accommodate some important challenges in visual scene understanding, i.e. semantic segmentation, texture classification, and object recognition. The aim is to accomplish specific tasks in some new high-dimensional spaces, covered by the statistical information of the datasets under study. Inspired by a new formulation of Fisher discriminant analysis, this thesis introduces some novel arrangements of well-known deep learning architectures, to achieve better performances on the targeted missions. The theoretical justifications are based upon a large body of experimental work, and consolidate the contribution of the proposed idea; Deep Fisher Discriminant Learning, to several challenges in visual scene understanding.