It has been shown that characteristic attributes of visually evoked potentials (VEPs) depend on the color and intensity of the stimulus. This may be helpful in different scenarios, for instance, to better understand (abnormal) physiological processes of color recognition or to optimize the visual stimulus of brain computer interfaces. Although previous works indicate that color discrimination is generally possible, novel methods for denoising and information extraction are needed to allow reliable classification of the stimulus shown to the subject. In this work, we investigated parametrized Hermite transformations that amplify subtle differences between VEPs induced by red and green lights. In order to compare the models, we built up our own dataset obtained from 9 individuals consisting of 1440 VEPs for classification. Then, we evaluated the discrimination power of the proposed Hermite features in a machine learning framework and compared the results to the state-of-the-art. We demonstrate a consistent and meaningful increase in classification accuracy due to the use of adaptive Hermite function based transformations.