Development of Feature Extractor for Visible-Light Iris Recognition Using Multi-task CNN
Tetsuya Honda, Hironobu Takano · 2021
The purpose of this study is to realize a visible-light iris authentication method with the same accuracy as personal authentication using near-infrared iris images. We proposed the feature extractor using the multi-task CNN (Convolutional Neural Network) and compared with the authentication accuracy of the single-task CNN. From the experimental results, the EER (Equal Error Rate) of multi-task CNN was up to 2.27% lower than that of single-task CNN. The authentication performance was improved by using the feature extractor of multi-task CNN.