Unconstrained Iris Segmentation Using Convolutional Neural Networks

Abstract

The extraction of consistent and identifiable features from an image of the human iris is known as iris recognition. Identifying which pixels belong to the iris, known as segmentation, is the first stage of iris recognition. Errors in segmentation propagate to later stages. Current segmentation approaches are tuned to specific environments.

We propose using a convolution neural network for iris segmentation. Our algorithm is accurate when trained on a single environment and tested on multiple environments. Our network builds on the Mask R-CNN framework (He et al. ICCV 2017). Our approach segments faster than previous approaches including the Mask R-CNN network.

Our network is accurate when trained on a single environment and tested with a different sensors (either visible light or near-infrared). Its accuracy degrades when trained with a visible light sensor and tested with a near-infrared sensor (and vice versa). A small amount of retraining of the visible light model (using a few samples from a near-infrared dataset) yields a tuned network accurate in both settings.

For training and testing, this work uses the Casia v4 Interval, Notre Dame 0405, Ubiris v2, and IITD datasets.

A Additional Related Work

As discussed in the introduction, segmentation algorithms can be classified into specialized, hybrid, and learning. We provide an overview of specialized and hybrid approaches below.

Specialized Approaches. The first generation of segmentation algorithms assumed that the iris and pupil are circular. Under this assumption circle/ellipse fitting techniques are used to find iris and pupil boundaries. Daugman’s seminal work uses an integro-differential operator for iris segmentation [9], in essence it tries to fit a circle over an iris exhaustively. Subsequent algorithms are based on integro-differential operators and the Hough transform [10, 17, 18]. Circle fitting via the Hough transform involves different regions voting for the best circle (which can be seen as an exhaustive search over a possible circle) [37]. New segmentation methods use different techniques to find candidate circles (for example, see Tan and Kumar [33]).

Methods that assume the iris and pupil are elliptical perform well in constrained environments where the individual faces the camera and actively participates in image collection. However, these methods perform poorly in unconstrained environments. As an example, in blurred iris images these algorithms fixate or diverge harshly, converging to edges besides the iris boundary. Substantial pre-processing and post-processing is necessary to use these methods in unconstrained environments.

Improvements to ellipse fitting have come in the form of generalizes structure tensors or GST [2]. GSTs find an iris specific complex circular pattern and convolve this pattern with the collected image to find the iris and pupil. The GST method allows the discovered region to only approximate a circle. Zhao and Kumar first use a total variation model to regularize local variations [38] and then feed this image into circular Hough transform. Interestingly, Zhao and Kumar process the lower half and the upper half of the iris separately.

Active contours [22] are used for general segmentation tasks. Instead of circle fitting, active contours find a high gradient between two sections in an image to indicate a boundary. Abdullah et al. show that with some iris specific modifications, active contours segment well [1]. While active contours do not assume the collected iris is circular they can still fixate on reflections and occlusions.

Hybrid Approaches. Machine learning techniques including neural networks have penetrated into fields that used specialized approaches. Many works segment the iris using a mix of general learning techniques and specialized iris techniques. The common approach is to use learning algorithms to do rough segmentation and post-process via specialized techniques to get the final segmented iris.

Proenca performed coarse iris segmentation using a neural network classifier and refined this segmentation via polynomial fitting [28]. Radman et al. use a HoG (Histogram of Gradients) as input features to train a support vector machine (SVM) [30]. This trained SVM is then used on new images to localize the iris. Subsequent iris segmentation is done by employing the Growcut algorithm [35] which labels pixels based on an initial guess. Adaboost [11] based eye detection is common for identifying the iris region inside a facial image, Joeng et al. [21] refined this technique by adding eyelid and occlusion removal algorithms. Tan and Kumar use Zernike moments as input features to train a neural network and SVM for coarse iris pixel classification [32].