CNN training for age group prediction in an illumination condition

CNN models trained with the given training dataset usually appear to show good accuracy in the testing with no added illumination. In this paper, we show that age group prediction in an added illumination condition results in a significant drop in accuracy. Our testing was performed on the color printed test photos captured through camera under office lighting condition. We also show the results of applying several possible training options to alleviate the accuracy drop such as using Grayscale or RGB images, severity of jitters of contrast and brightness on images, and the augmentation of training data.