WECIA Graph: Visualization of Classification Performance Dependency on Grayscale Conversion Setting

Abstract

Grayscale conversion is a popular operation performed within image pre-processing of many computer vision systems, including systems aimed at generic object categorization. The grayscale conversion is a lossy operation. As such, it can significantly influence performance of the systems. For generic object categorization tasks, a weighted means grayscale conversion proved to be appropriate. It allows full use of the grayscale conversion potential due to weighting coefficients introduced by this conversion method. To reach a desired performance of an object categorization system, the weighting coefficients must be optimally setup. We demonstrate that a search for an optimal setting of the system must be carried out in a cooperation with an expert. To simplify the expert involvement in the optimization process, we propose a WEighting Coefficients Impact Assessment (WECIA) graph. The WECIA graph displays dependence of classification performance on setting of the weighting coefficients for one particular setting of remaining adjustable parameters. We point out a fact that an expert analysis of the dependence using the WECIA graph allows identification of settings leading to undesirable performance of an assessed system.