Object-specific Feature Extraction via Markov Random Fields Derived from 0-order Sigma-tree Segmentations

Sigma-Trees associated with residual vector quantization (RVQ) has been used for image-driven data mining to detect features and objects in a digital image with a degree of success. RVQ methods based on σ-tree structures have been designed to implement successive refinement of information for image segmentation. In such implementations, RVQ based novel methods are devised for pixel-block mining, pattern similarity scoring, class label assignments and attribute mining (Barnes, 2007a). Direct sum σ-tree structures are used for near-neighbor similarity scoring. The variable bit-plane data representations produced by σ-tree structures not only provides an approach for image content segmentation and a structure for formulation of Bayesian classification, but also offers a solution to the challenge of high computational costs involved in pixel-block similarity searching. Such σ-tree based multi-stage RVQ classifiers have yielded promising image-content segmentation/classification yielding fine-grained features extraction. This ability to produce fine-grained features has been exploited in object detection applications. However, in the context of object identification the methods have been applied heuristically on single stages of the multi-stage σ-tree based direct sum successive refinement data representation. As a result, object detection with optimal rejection of false alarm is not guaranteed. Gibbs random field (GRF), also known as Markov random field (MRF), provides a joint probabilistic framework to model the object identification task in digital images. As a result, the image segmentation task can be solved optimally in the maximum aposteriori probabilistic (MAP) sense. Thus, the advantages of the σ-tree based RVQ classifier to provide fine-grained feature extractions for object of interest can be exploited with an MRF-based model of the object. This paper demonstrates the use of MRF on a 0 order output of the σ-tree based RVQ for the purpose of object detection.