We have applied CNN Ensemble with Rejection to detect product defects. Our experimental results have shown that these two CNN models can detect product defects very accurately, especial when they are used as a member of an ensemble. Among them, GoogLeNet can make more accurate judgments and reduce the number of rejections effectively.