Abstract
Face recognition has dramatically improved its performance with the advent of deep learning, especially Convolutional Neural Networks (CNNs), while they have raised a new issue: the difficulty in interpreting the results. The question is what CNN looks at in a face image to identify a person. To answer this question, this paper presents a simple and novel analysis of deep face recognition based on facial parts. We evaluate the recognition accuracy of face images with specific regions masked using face segmentation labels. Our analysis clarifies what CNNs really need in face images for face recognition. The paper concludes with an application of face recognition models to general visualization methods and the problems contained in some classical face image datasets.