Parameter Study and Optimization of a Color-Based Object Classification System

Abstract

Typical computer vision systems usually include a set of components such as a preprocessor, a feature extractor, and a classifier that together represent an image processing pipeline. For each component there are different operators available. Each operator has a different number of parameters with individual parameter domains. The challenge in developing a computer vision system is the optimal choice of the available operators and their parameters to construct the appropriate pipeline for the problem at hand. The task of finding the optimal combination and setting depends strongly on the definition of the term optimal. Optimality can reach from minimal computational time to maximal recognition rate of a system. Using the example of the color-based object classification system, this contribution presents a comprehensive approach for finding an optimal system by defining the required image processing pipeline, defining the optimization problem for the classification and improving the optimization by taking parameter studies into consideration. This unique approach produces a color-based classification system with an illuminant independent structure.