Iris Segmentation with Multi-task LearningKREHO, ADNAN (Avtor) Štruc, Vitomir (Mentor) iris segmentationimage colorizationmulti-task learningUNet architectureOcular biometrics are heavily utilized today, especially in security systems. There are multiple ocular features that can be used for efficient recognition and identification: from the iris to the periocular region. In this thesis, we adress the problem of iris segmentation, which is the first step of an iris recognition system that affects the performance of all downstream tasks. We design a novel approach that utilizes multi-task learning in order to build a robust model capable of segmenting the iris area from a gray-scale image. Specifically, the developed models aim at simultaneously segmenting the iris area and at performing image colorization of the input gray-scale ocular image. By learning to solve these tasks jointly, the model leverages task correlations and domain-specific knowledge to improve performance of both tasks. We consider two architectures in our experiments, UNet and UNet2, which are similar in structure. However, UNet2 has half the number of feature channels in every convolutional layer except the last one. Because of this, UNet2 has a significantly lower number of trainable parameters. We experiment with two models to test which architecture is more efficient with respect to the iris segmentation task, and which is faster to train. Also, for multi-task learning, we test how different sets of loss weights affect the quality of iris segmentation. The models are trained, evaluated and tested on the MOBIUS and UBIRIS.v2 datasets. The developed models are also compared to the IrisParseNet networks trained on UBIRIS.v2 and MICHE datasets. We note at this point that the IrisParseNet networks are the current state-of-the-art in solving the iris segmentation task. Based on the evaluation results we conclude that models utilizing multi-task learning perform better than models utilizing the single-task learning approach. We also show that models, that use a higher loss weight for the iris segmentation task, prove to be better-performing, which is expected given that most of the attention during training is given to the iris segmentation task. Additionally, our models perform better than the IrisParseNet models in most of our experiments. The observed results are very encouraging for future work on this topic, and point to multi-task learning as the go-to approach for solving the iris segmentation task.20212021-11-02 12:20:02Magistrsko delo/naloga132753sl