Random forest analysis and its applications in computer vision

Abstract

Random forest is a well-known machine-learning technique, which has been widely employed in machine-learning and computer-vision applications, including classification, clustering and regression, due to its strong generalization power and high efficiency on both its training and inference stages. A random-forest based model consists of an ensemble of binary trees, which build a forest. Every of the trees can make its own decision independently as an individual expert, and the final prediction is the result summarized from all the trees. Although the vanilla version of random forest has been largely used as a standard tool integrated in lots of software packages, random forest is still a research area where some aspects can be improved, such as the criterion in the split nodes, the clustering algorithm in the leaf nodes, and the feature engineering in the whole process. In this thesis, we address these problems and have proposed some novel ideas which make remarkable improvements on the random-forest-based models. Firstly, we propose a random-forest-based, cascaded regression model for face alignment, by designing a novel locally lightweight feature, namely intimacy definition feature (IDF), which can achieve high speed and stable accuracy for applications where hardware resources are limited, such as mobile devices. Secondly, we present a more accurate face alignment algorithm by combining IDF and the classic constrained local model (CLM) paradigm into a joint framework. After studying the impact of feature distribution in a random forest, we propose a novel random-forest scheme, namely Joint Maximum Purity Forest (JMPF), for classification, clustering, regression and image super-resolution, where the original feature space is transformed into a compactly pre-clustered feature space, via a trained rotation matrix. Finally, we present a novel feature-augmented random forest (FARF) for image super-resolution, where we have studied feature engineering in the whole process of random forest and extended some work from JMPF. All the algorithms proposed in this thesis have been evaluated and compared to existing state-of-the-art methods. Experimental results and analyses show that our algorithms can achieve stable and promising performances.