The effect of attribute normalization factors in attribute distance weighted average

Abstract

Attribute distance weighted average (ADWA) is a new filtering paradigm, which can progressively alleviate the denoising contradiction between the noise removal and feature preservation by introducing new attributes. As the key control parameters in ADWA, the attribute normalization factors play an important role in the final filtering result. An in-depth study is necessary to exam the effect the attribute normalization factors have on the filtering performance and the rules they follow, which can then serve as a guide for the determination and optimization of attribute normalization factors. For this purpose, the three attributes of a signal, “Location,” “Value,” and “Gradient,” are studied as an example in this paper. Experimental results indicate that the normalization factors directly determine the strength of the effect the corresponding attributes have on the filtering result. If the normalization factor increases, ADWA’s ability in noise removal becomes stronger and meanwhile its ability in feature preservation becomes weaker. Therefore, the denoising contradiction still exists for ADWA of a specific attribute rank. However, since different attributes contribute to the filtering performance independently in different regions of a signal, the denoising contradiction can be further alleviated by introducing new attributes, and thus a more satisfactory outcome can be obtained.