Hierarchical Image Segmentation by Structural Content

Abstract

Image quality loss resulting from artifacts depends on the nature and strength of the artifacts as well as the context or background in which they occur. In order to include the impact of image context in assessing artifact contribution to quality loss, regions must first be classified into general categories that have distinct effects on the subjective impact of the particular artifact. These effects can then be quantified to scale the artifact in a perceptually meaningful way. This paper formulates general context categories, develops automatic image region classifiers, and evaluates the classifier performance using images containing multiple categories. Linear classifiers are designed to identify three main classes which include random, textured, and transient regions. Features for identifying these areas over regions at multiple resolutions are based on the optical density histogram (ODH), the cortex transform, and the cooccurrence matrix. It was found that selecting features from the ODH and cortex transform provides classification results in agreement with human assessment, and performances comparable to those of classifiers using cooccurrence matrix features. Experiments to assess performance show misclassification rates ranging from 3.3% for the lowest resolutions to 32.2% at highest. This paper also presents a hierarchical classification algorithm that combines classifiers operating at multiple resolutions and achieves an overall misclassification rate as low as 4.8%.