Statistical feature selection and extraction for video and image segmentation

Abstract

Scope and Method of Study: The purpose of this study was to develop statistical feature selection and extraction methods for video and image segmentation, which partition a video or image into non-overlap and meaningful objects or regions. It is a fundamental step towards content-based visual information analysis. Visual data segmentation is a difficult task due to the various definitions of meaningful entities, as well as their complex properties and behaviors. Generally, visual data segmentation is a pattern recognition problem, where feature selection/extraction and data classifier design are two key components. Pixel intensity, color, time, texture, spatial location, shape, motion information, etc., are most frequently used features for visual data representation. Since not all of features are representative regarding visual data, and have significant contribution to the data classification, feature selection and/or extraction are necessary to select or generate salient features for data classifier. Statistical machine learning methods play important roles in developing data classifiers. In this report, both parametric and nonparametric machine learning methods are studied under three specific applications: video and image segmentation, as well as remote sensing data analysis.

Findings and Conclusions: For various visual data segmentation tasks, key-frame extraction in video segmentation, WDHMM likelihood computation, decision tree training, and support vector learning are studied for feature selection and/or extraction and segmentation. Simulations on both synthetic and real data show that the proposed methods can provide accurate and robust segmentation results, as well as representative and discriminative features sets. This work can further inspire our studies towards the real applications. In these applications, we are able to obtain state-of-the-art or promising results as well as efficient algorithms.