A Neural Markovian Multiresolution Image Labeling Algorithm

TL;DR

This paper evaluates a hierarchical image labeling algorithm that merges regions using Markov random fields, producing a sequence of partitions that can aid higher-level vision tasks, with promising subjective and objective results.

Contribution

It introduces a multiresolution, hierarchical image labeling algorithm based on Markov random fields that outputs a sequence of partitions for enhanced vision analysis.

Findings

Very good results on subjective criteria

Excellent performance on objective criteria

Applicable to 1D, 2D, and 3D signals

Abstract

This paper describes the results of formally evaluating the MCV (Markov concurrent vision) image labeling algorithm which is a (semi-) hierarchical algorithm commencing with a partition made up of single pixel regions and merging regions or subsets of regions using a Markov random field (MRF) image model. It is an example of a general approach to computer vision called concurrent vision in which the operations of image segmentation and image classification are carried out concurrently. While many image labeling algorithms output a single partition, or segmentation, the MCV algorithm outputs a sequence of partitions and this more elaborate structure may provide information that is valuable for higher level vision systems. With certain types of MRF the component of the system for image evaluation can be implemented as a hardwired feed forward neural network. While being applicable to images (i.e. 2D signals), the algorithm is equally applicable to 1D signals (e.g. speech) or 3D signals (e.g. video sequences) (though its performance in such domains remains to be tested). The algorithm is assessed using subjective and objective criteria with very good results.