Although CNN-based classifiers have been successfully applied to object recognition, their performance is not consistent. In particular, when a CNN-based classifier is applied to a new dataset, the performance can substantially deteriorate. Furthermore, classification accuracy for certain classes can be very low. In many cases, the poor performance of ill-performing classes is due to biased training samples, which fail to represent the general coverage of the ill-performing classes. In this paper, we explore how to enhance the training samples of such ill-performing classes based on coverage optimization measures. Experimental results show some promising results.