Robust Image Labeling Using Conditional Random Fields

Abstract

Object recognition has become a central topic in computer vision applications such as image search, robotics and vehicle safety systems. However, it is a challenging task due to the limited discriminative power of low-level visual features in describing the considerably diverse range of high-level visual semantics of objects. Semantic gap between low-level visual features and high-level concepts are a bottleneck in most systems. New content analysis models need to be developed to bridge the semantic gap. In this thesis, algorithms based on conditional random fields (CRF) from the class of probabilistic graphical models are developed to tackle the problem of multiclass image labeling for object recognition. Image labeling assigns a specific semantic category from a predefined set of object classes to each pixel in the image. By well capturing spatial interactions of visual concepts, CRF modeling has proved to be a successful tool for image labeling. This thesis proposes novel approaches to empowering the CRF modeling for robust image labeling. Our primary contributions are twofold. To better represent feature distributions of CRF potentials, new feature functions based on generalized Gaussian mixture models (GGMM) are designed and their efficacy is investigated. Due to its shape parameter, GGMM can provide a proper fit to multi-modal and skewed distribution of data in nature images. The new model proves more successful than Gaussian and Laplacian mixture models. It also outperforms a deep neural network model on Corel imageset by 1% accuracy. Further in this thesis, we apply scene level contextual information to integrate global visual semantics of the image with pixel-wise dense inference of fully-connected CRF to preserve small objects of foreground classes and to make dense inference robust to initial misclassifications of the unary classifier. Proposed inference algorithm factorizes the joint probability of labeling configuration and image scene type to obtain prediction update equations for labeling individual image pixels and also the overall scene type of the image. The proposed context-based dense CRF model outperforms conventional dense CRF model by about 2% in terms of labeling accuracy on MSRC imageset and by 4% on SIFT Flow imageset. Also, the proposed model obtains the highest scene classification rate of 86% on MSRC dataset.