Performance Improvement of 3D-CNN for Blink Types Classification by Data Augmentation

Abstract

When developing a blink input interface, conscious (voluntary) and natural (involuntary) blink types must be automatically classified. We previously proposed a method for blink type classification using a 3D convolutional neural network (3D CNN). This CNN model outputs a predicted probability that determines three classes: “voluntary blinking,” “involuntary blinking,” and “not blinking” from a periocular image sequence. Previously, we found that the bias of the eye position in the input image is a factor that reduces the classification accuracy. To address this problem, we employ data augmentation with a shifting 5 or 10 pixels in the horizontal and/or vertical directions.