The eye is the most important part of the facial features. Studying the feature point location and shape classification of the eye plays an important role in face recognition. This paper proposes a human eye classification method based on cascaded convolutional neural network and semantic features. Three levels of cascaded convolutional neural networks were used to detect and locate 106 feature points from rough to meticulous, among which 20 feature points can accurately locate the eyes. Based on these 20 feature points, the shape of the eye is modeled and three shape parameters defining the eye shape are defined. These three parameters are segmented to achieve the purpose of eye classification according to different semantic descriptions of each interval. Experimental results show that this method is accurate and can achieve good eye classification effect.