Table 1.

Comparisons between the proposed and existing methods for iris segmentation.

Type	Methods	Strength	Weakness
Iris circular boundary detection without eyelid and eyelash detection	Iris localization by circular HT [17,22,24]	These methods show a good estimation of the iris region in ideal cases	These types of methods are not very accurate for non-ideal cases or visible light environments
	Integro-differential operator [18]
	Iris localization by gradient on iris-sclera boundary points [19]	A new idea to use a gradient to locate iris boundary	The gradient is affected by eyelashes and true iris boundary is not found
	The two-stage method with circular moving window [20]	Pupil based on dark color approximated simply by probability	Calculating gradient in a search way is time-consuming
	Radial suppression-based edge detection and thresholding [21]	Radial suppression makes the case simpler for the iris edges	In non-ideal cases, the edges are not fine to estimate the boundaries
	Adaptive thresholding and first derivative-based iris localization [23]	Simple way to obtain the boundary based on the gray level in ideal cases	One threshold cannot guarantee good results in all cases
Iris circular boundary detection with eyelid and eyelash detection	Two-circular edge detector assisted with AdaBoost eye detector [26]	Closed eye, eyelash and eyelid detection is executed to reduce error	The method is affected by pupil/eyelid detection error
	Curve fitting and color clustering [27]	Upper and lower eyelid detections are performed to reduce the error	The empirical threshold is set for eyelid and eyelash detection, and still true boundary is not found
	Graph-cut-based approach for iris segmentation [28]	Eyelashes are removed using Markov random field to reduce error	A separate method for each eyelash, pupil, iris detection is time-consuming
	Rotated ellipse fitting method combined with occlusion detection [29]	Ellipse fitting gives a good approximation for the iris with reduced error	Still, the iris and other boundaries are considered as circular
	Three model fusion-based method assisted with Daugman’s method [30]	Simple integral derivative as a base for iris boundaries is a quite simple way	High-score fitting is sensitive in ideal cases, and can be disturbed by similar RGB pixels in the image
Active contour-based methods	Geodesic active contours, Chan–Vese and new pressure force active contours [31,32,33,34]	These methods iteratively approximate the true boundaries in non-ideal situations	In these methods, many iterations are required for accuracy, which takes much processing time
Region growing and watershed methods	Region growing with integro-differential constellation [35]	Both iris and non-iris regions are identified along with reflection removal to reduce error	The rough boundary is found first and then a boundary refinement process is performed separately
	Region growing with binary integrated intensity curve-based method [36]	Eyelash and eyelid detection is performed along with iris segmentation	The region growing process starts with the pupil circle, so the case of visible light images where the pupil is not clear can cause errors
	Watershed BIRD with seed selection [37]	Limbus boundary detection is performed to separate sclera, eyelashes, and eyelid pixels from iris	Watershed transform shares the disadvantage of over-segmentation, so circle fitting is used further
Deep-learning-based methods	HCNNs and MFCNs [44]	This approach shows the lower error than existing methods for non-ideal cases	The similar parts to iris regions can be incorrectly detected as iris points
	Two-stage iris segmentation method using deep learning and modified HT [45]	Better accuracy due to CNN, which is just applied inside the ROI defined in the first stage	Millions of 21 × 21 images are needed for CNN training and pre-processing required to improve the image
	IrisDenseNet for iris segmentation (Proposed Method)	Accurately find the iris boundaries without pre-processing with better information gradient flow. With robustness for high-frequency areas such as eyelashes and ghost regions	Due to dense connectivity, the mini-batch size should be kept low owing to more time required for training