Deep Learning-Based Methodologies

Deep learning-based approaches are the primary source of majority of state-of-the-art solutions. In segmentation, researchers have come up with numerous modifications of simple auto-encoder-based CNNs. UNet, as mentioned earlier, has now become the basis network for segmentation tasks. In the iris segmentation, researchers shifted towards the deep neural network-based approaches gradually as the availability of the iris databases increased, and classical methods were challenged with increasing non-ideality in the iris images. CNN remove pre-/post-processing steps. CNN models were introduced to increase the accuracy of segmentation masks generated over the non-ideal iris images.

Liu et al. [63] developed two modalities with CNN, where a multi-scale CNN along with hierarchical CNN was deployed to detect the iris boundaries in non-ideal cases. In Ref. [59], the authors fed the input iris images to a series of four dense convolution blocks; feature maps extracted from blocks fused with a weighted sum gave coarse as well as fine features to produce the required segmentation mask finally. The authors of Ref. [71] trained two different CNN architectures, which were derived from networks such as Faster RCNN [82] and SSD [64] which localise the circular region of the pupil along with iris. Similarly in Ref. [11], the authors combined two existing CNN networks: the DenseNet [44] and the SegNet [12] for iris segmentation. Another such work which incorporates the features of one CNN architecture into another is Ref. [10]; they developed a CNN model with residual connections in SegNet which allowed the authors to develop deeper network while reducing the chances of vanishing gradient. In Ref. [63], the authors developed two CNNs, the first one was based on hierarchical CNNs and the second one was based on multi-scale CNNs. The authors of Ref. [14] used the data augmentation approach along with the CNN model to virtually increase the non-ideal iris database for segmentation.

In Ref. [104], the authors took the non-ideality to the next level, where the iris images were taken using the mobile images. They developed a lightweight deep CNN as a complete end-to-end segmentation method. Followed by the previous work, Wang et al., in their work [103], came up with a multi-task CNN architecture that also incorporates the attention module for the iris segmentation and boundary localisation. Similarly, in Ref. [54], the authors developed EyeNet, an attention-based CNN for the eye region segmentation.

While numerous works have been published on the iris segmentation with CNN both over ideal and non-ideal images, more or less, they propose a few' new additional features or modules towards the existing models. Challenges that the deep learning solutions try to solve are a variety of textural complexity and the shape of iris from person to person, non-ideality such as distortion, non-regular illumination, motion blur, digital noise, and poor image quality. Algorithms are not robust enough to extract the ROI for segmentation. One such work that handles the majority of the above issues while giving state-of-the-art segmentation maps is Ref. [53], wherein the authors proposed a three-stage trained novel deep CNN architecture for the nonideal iris segmentation. In addition to novel models, they used a combination of multiple loss functions to give precise segmentation maps. Table 12.3 compares some selected deep learning-based methods.