Efficient and robust eye images iris segmentation using a lightweight U-net convolutional network

Abstract

The paper presents an efficient lightweight U-net convolutional neural network (CNN) architecture that can be used for iris segmentation in eye images. The novelty of the proposed method consists of model downscaling for efficiency, while maintaining high iris segmentation accuracy. The network is validated on different resolution and quality images from five standard open source benchmarks: BioSec, CasiaI4, CasiaT4, IITD, and UBIRIS. The efficient U-net architecture consists of 36 layers and uses 148 k parameters, value order of magnitude lower than other existing networks used for similar applications. This also leads to a much lower training time and eye image segmentation time (< 1 ms per image on Xeon CPU). The iris segmentation results obtained were state-of-the-art in terms of standard nice1, F1 and mIoU accuracy measures on all the analyzed datasets. Whilst some differences can be observed for these measurements between datasets, the lowest values for F1 and mIoU parameters obtained were 96.14% and 92.56%, respectively, on UBIRIS dataset, and for nice1 parameter 0.38 on CasiaT4. The best results were obtained on CasiaI4 dataset with F1 = 98.61%, mIoU = 97.26%, and nice1 = 0.78.