Support Vector Random Fields for Spatial Classification

In this paper we propose Support Vector Random Fields (SVRFs), an extension of Support Vector Machines (SVMs) that explicitly models spatial correlations in multi-dimensional data. SVRFs are derived as Conditional Random Fields that take advantage of the generalization properties of SVMs. We also propose improvements to computing posterior probability distributions from SVMs, and present a local-consistency potential measure that encourages spatial continuity. SVRFs can be efficiently trained, converge quickly during inference, and can be trivially augmented with kernel functions. SVRFs are more robust to class imbalance than Discriminative Random Fields (DRFs), and are more accurate near edges. Our results on synthetic data and a real-world tumor detection task show the superiority of SVRFs over both SVMs and DRFs.