Noise reduction for pattern-matching algorithms

Abstract

Many of the data analysis algorithms that base their analysis on pattern occurrences tend to use objective assessment measures at one point or another. In many cases, especially in multimedia research, these objective measures were originally developed for the purpose of mimicking subjective assessments to automate the assessment pipeline. Using such measures is understandable when the user is a human subject but becomes arguable when there is an intermediate user in the form of an analysis algorithm. We present here a multidimensional noise reduction scheme that cleans the data from the perspective of the algorithmic system. The proposed scheme is then applied to applications of image coding and content-based image retrieval. Although the noise reduction adversely affects the objective scales, we show that it actually enhances the performance of the analysis algorithm. For instance, the percentage retrieval precision of tiger images was 3.5-fold better than the non-enhanced system. This precision enhancement is accompanied by a 55% reduction in retrieval time on average and further reduction in space costs.