Augmentation in Neural Network Training for Person Identification by Iris Images

摘要

Convolutional neural networks are at the heart of most modern solutions for various computer vision problems. The paper compares the effectiveness of using several convolution neural network architectures, namely, DenseNet, EfficientNet, and Inception V3 for solving the problem of person identification by eye's iris images. Besides, we evaluate two types of augmentation for convolutional neural network training, which can improve the classification accuracy. The first type is to apply a random segmentation mask from the bank of segmentation masks to training images, and the second type is to crop images randomly. All experiments were conducted using the MMU Iris Database, which contains 450 images for 45 classes. Experimental studies have shown that the application of the approaches proposed in work is a somewhat effective method for solving the identification problem. The best quality classification of the iris was obtained using convolutional neural network architecture DenseNet and augmentation of the first type.