Review of iris segmentation and recognition using deep learning to improve biometric application

1 Introduction

The technology of biometric recognition is becoming increasingly crucial for individual identification, with a wide range of applications that include authentication of mobile devices, as well as security and surveillance. Iris recognition (IR) is considered to be a highly dependable biometric technology that utilizes distinct iris patterns to accurately authenticate individuals [1]. The precision of IR is contingent upon the precision of iris segmentation, which involves the isolation of the iris area from the remaining portion of the ocular image. The discipline of biometrics pertains to the process of verifying and validating the identity of individuals by means of their distinct biological and behavioral traits [2]. Numerous studies concentrate on contemporary developments in identifying the most essential biometric features, with particular attention to techniques for iris segmentation and recognition processing. Biometric systems are assessed based on their design, operational procedures, and performance metrics, as stated in the study of Nachar and Inaty [3]. Biometric identification systems utilize measurable and observable physical and behavioral traits to authenticate an individual’s identity. The aforementioned characteristics encompass inherent physiological features, namely IR, fingerprints, facial features, retinas, veins, vocal patterns, and hand geometries [4].

The utilization of IR is widely acknowledged as a highly secure and precise biometric technique, primarily attributed to the iris’s capacity to retain diverse phase information, encompassing an estimated 249 degrees of freedom [5]. The iris is an annular, chromatic structure located within the ocular globe that modulates the quantity of light that penetrates the retina, and its demarcation encompasses the central aperture known as the pupil and the outermost layer called the sclera. According to research, even in the scenario of monozygotic twins, every pair of eyes will exhibit a marginally distinct iris pattern that persists over the course of their lifespan [6]. The distinct visual characteristics of the iris are attributed to its distinguishing features, including cellars, radial grooves, threads, pigment frills, spots, stripes, and longitudinal ligaments [7].

Despite the increasing popularity of biometric systems, their present state precludes their universal applicability. Consequently, it is imperative to improve the security and adaptability of biometric systems to ensure their widespread accessibility [8]. The property of high distinguishability is considered to be a fundamental aspect of dependable biometric systems, as noted in the study of Kagawade and Angadi [9]. The attainment of a high level of biometric differentiation certainty is of utmost significance, especially in the context of biometrics employed for monitoring human health or in security applications with high stakes. In the absence of it, there is a potential for the removal of the items from the shelves and consequent harm to individuals or property damage [10].

The attribute of persistence is a vital component of biometrics, as demonstrated by the precise and dependable observation and acquisition of fundamental physical characteristics of users through iris segmentation and recognition [11]. The iris is considered to be a highly reliable biometric characteristic due to its distinct and complex patterns. According to research, irises exhibit distinct characteristics even in cases of fraternal twins or the left and right eyes of an individual [12]. The utilization of IR as a biometric method is deemed highly secure and precise, thus presenting a promising avenue for future research and development.

Several pre-trained models in machine vision, including VGGNet, ResNet, DenseNet, and NTU, have exhibited robust performance in classification and identification tasks, as evidenced by prior research [13]. The utilization of transfer learning could potentially prove advantageous in the training of models based on IR. Figure 1 depicts conventional techniques utilized for the detection of a viable iris. In the realm of iris-liveness detection, deep learning (DL) models are the prevailing approach for constructing models and performing classification tasks. Although DL has emerged as a prominent trend in this field, it is improbable that it will entirely supplant current methodologies in the immediate future [14].

Figure 1

Iris real-time intrusion detecting methods [14].

In order to attain an optimal biometric, certain fundamental aspects of physical appearance must remain constant and immutable, as stated in the study of Boyd et al. [15]. Nonetheless, the capacity to be gathered is also a pivotal aspect of an optimal biometric. This statement suggests that the biometric system must effectively and efficiently authenticate user characteristics and data. Furthermore, the formidable obstacle of addressing the exorbitant production cost associated with conventional biometric technology must be tackled.

The main contributions of this study are sorted as follows:

1. Examination and evaluation of biometric applications that utilize DL techniques for iris segmentation and recognition.

2. Improving accuracy in the areas of authentication and detection with different research directions.

3. The discourse surrounding the fundamental characteristics that a proficient biometric ought to exhibit encompassing attributes such as security, confidentiality, adaptability, and superior discernment.

4. The main motivation for iris segmentation and recognition using DL in the context of biometric applications is significantly less unauthorized access is possible, and real-time analysis of enormous amounts of biometric data is possible.

5. It offers significant perspectives and tools for scholars and professionals operating within the realm of biometrics. Table 1 shows the used methodology to solve the existing issues.

In Section 2, a comprehensive overview of IR and segmentation techniques is provided, encompassing both conventional and DL-based approaches. In Section 3, extant literature pertaining to DL-based iris identification technologies and resources is examined. Additionally, evaluation methods, including but not limited to reliability, specificity, recollection, and F-score, are discussed. Section 4 delves into contemporary advancements in the analysis of iris images, edge detection techniques, and categorization methods. Section 5 provides a comprehensive assessment of IR techniques that are based on DL, along with an analysis of the difficulties that are inherent to these methods. Section 6 provides a conclusive analysis of the potential of DL in the context of biometric identification. The section also offers recommendations for future research aimed at addressing the current limitations of IR systems and enhancing their performance.

2 Enhancing biometric applications through DL-based iris segmentation and recognition accuracy

IR and retinal scanning are considered to be highly dependable and precise biometric modalities utilized for the purposes of authentication and verification. Iris segmentation is a biometric modality that shares similarities with fingerprint recognition and facial identification. Law enforcement authorities have the capability to conduct a comparative analysis of iris images of individuals suspected of criminal activity against a pre-existing database of images for the purpose of ascertaining or confirming their identity. The high accuracy of IR can be attributed to the distinctiveness of iris patterns, which results in a negligible occurrence of false positives. Consequently, the process of iris segmentation is deemed crucial in enhancing the precision of IR. Compared to other biometric applications, the DL method exhibits superior performance in image analysis and matching, with more favorable tradeoff curves between errors of refusal and acceptance, as shown in Table 2.

Table 2 outlines the various DL models employed in the training process, along with the corresponding dataset utilized. An iris identification methodology entails analyzing a representation of the data that is devoid of dimensionality and constructing a feature vector that is positioned at the apex of a framework. The study presented in Table 3 investigated the hypothesis that iris-specific feature extractors would exhibit superior performance compared to generic models. This was accomplished by examining six weight configurations for the commonly utilized ResNet50 architecture. The authors observed that the utilization of CASIA-IrisV3 yielded comprehensive characteristics for iris authentication, which were found to be effective in biometric identification. The attainment of high memory space, training samples, training time, features, performance, and dependability poses several challenges in current methodologies.

3 Biometric authentication using iris segmentation

The process of IR, also referred to as iris scanning, involves the capture of high-contrast images of an individual’s irises using visible and near-infrared light for subsequent analysis. Similar to fingerprints and facial recognition, biometrics encompasses it. Despite the higher accuracy rate of iris scanning in comparison to fingerprint scanning, it is imperative to establish a direct visual connection between the scanner and the eye. While IR is considered more secure than fingerprint identification, the latter is deemed more practical due to its ability to function without requiring a direct visual line of sight. The biometric technology community has shown significant interest in DL-based IR and segmentation due to its high level of dependability. IR is a biometric verification technique that leverages the distinct patterns of dark and light present in an individual’s irises to establish their identity. A computerized biometric identification process called IR analyses distinctive patterns in the ring-shaped area around each eye’s pupil. With very low false match rates, it is an exceptionally accurate and trustworthy identification approach. However, it has significant drawbacks, such as the need for specialized equipment [31].

The main advantages of combining IR with other biometric modalities are as follows:

1. High accuracy.

2. Speed and ease-of-use.

3. Unique and stable.

Besides, the disadvantages are as follows:

1. Costly infrastructure.

2. Privacy concerns.

3. Environmental constraints.

4. Limited user acceptance.

Furthermore, the system is capable of accessing the biometric dataset, as shown in Table 4.

Table 5 presents a comparative analysis of diverse DL methodologies employed in the context of IR and segmentation, utilizing a range of iris datasets. The present study offers a comprehensive analysis of the efficacy of various DL models in the context of iris biometric applications. The literature review of the article also examines diverse state-of-the-art implementations of CNNs in the field of biometrics. The researchers conducted a comparative analysis of various methodologies utilizing several datasets, namely CASIA interval v3, CASIA interval v1, NTU, UniNet.V2, IITD, Alex Net, Res Net, and Dense Net. A salient characteristic of IR utilizing CNNs is the identification of users through the analysis of distinct patterns. The system has the capability to recognize an individual by utilizing the distinct characteristics of their iris, which are obtained during the phase of image acquisition [41].

The utilization of gravitational motion analysis is a potential method for attaining iris verification. The methodology entails the examination of alterations in the iris’s position over a period of time, thereby furnishing supplementary data for biometric authentication. The advanced filtering capabilities of CNN are emphasized as a notable advantage in the context of IR and segmentation. These techniques facilitate the retrieval of significant latent features from intricate data sets with high dimensions, which would otherwise be challenging to extract using traditional biometric and measurement methodologies. It is imperative to acknowledge that unanticipated and insoluble challenges may arise during the process of data collection. In the context of IR, potential challenges may encompass variables such as variations in illumination, obstructions arising from eyelids or eyelashes, or reflections originating from eyewear. Notwithstanding these obstacles, CNNs are demonstrating remarkable efficacy in the realm of biometric authentication [42].

4 Biometric adaptability using a DL technique

DL is a highly effective machine learning methodology that utilizes neural networks consisting of multiple layers. It has found extensive application in diverse biometric domains. The popularity of utilizing biometric applications stems from their exceptional results, ability to withstand noise and variability, and versatility in accommodating diverse tasks and their unique characteristics. The system under consideration exhibits the capacity to adjust to diverse samples of frontal faces in the context of fingerprint, iris, and face recognition and segmentation. Furthermore, in biometric applications, DL models demonstrate enhanced adaptability through the utilization of inter-session and inter-task data, as shown in Table 6.

Table 7 presents a comprehensive summary of the performance metrics, encompassing accuracy, precision, and F1 score, of the DL methodologies examined in this review. The findings indicate that the incorporation of sophisticated technologies for biometric adaptation results in enhanced precision in diverse biometric identification assignments.

5 Iris and fingerprint fusion for multi-modal biometrics

Multi-modal biometric systems have become increasingly popular in recent years owing to their superior accuracy and reliability in comparison to single-modal systems. The present investigation employed several score normalization techniques, namely hyperbolic tangent, Z-score, and min–max. These techniques are implemented to guarantee that the scores fall within a designated range and are capable of being compared. It employs various score fusion techniques, namely user weighting, minimum scores, maximum scores, and simple sum. The aforementioned techniques are utilized to amalgamate the standardized scores derived from the iris and fingerprint information, resulting in a solitary integrated score. Subsequently, the amalgamated scores are employed to categorize an unidentifiable user as either an imposter or an authentic. The initial phase of fingerprint processing encompasses three distinct stages, namely image enhancement, identification of regions of interest (ROIs), and extraction of relevant features [49]. This procedure guarantees the high quality of the fingerprint data and exclusively employs pertinent characteristics in the amalgamation procedure, as shown in Table 8.

Table 9 provides a comparative evaluation of conventional and DL methodologies for IR and fingerprint fusion in multi-modal biometric systems. The system’s performance evaluation is conducted through the evaluation of the classifiers’ precision, recall, and F1 score on IR, fingerprint recognition, and the fusion of iris and fingerprint for biometric authentication. The performance of a classifier in accurately categorizing an input image based on the provided label is evaluated using metrics such as accuracy, F1 score, precision, and specificity [54]. The findings indicate that the utilization of DL-based methodology exhibits superior accuracy compared to the conventional approach, thereby highlighting the potential of DL techniques for multi-modal biometric applications.

The term “classification sensitivity” pertains to the ability of a classifier to accurately detect positive cases, whereas “specificity” pertains to the ability of the same classifier to accurately detect negative cases. Multiple supplementary texture-generating techniques are integrated to enhance the efficacy of the feature extraction methodology. The VGG16 classifier employs the kernel function to facilitate IR. Additionally, many classifiers are used for biometrics and face detection and other methods [54,55]. There are different limitations in iris segmentation and recognition using DL to improve biometric applications exhibits enhanced dependability and a higher degree of error elimination, the utilization of line feature-based fingerprint recognition presents various benefits in comparison to conventional minutiae-based features, reduced template sizes, decreased memory consumption, expedited matching time, and enhanced resilience to image misrepresentation.

6 Conclusion

The objective of biometrics is to accurately and distinctively authenticate an individual by utilizing one or multiple physiological and/or behavioral characteristics. This article presented a comprehensive literature review of biometric systems that possess significant potential for expansion and advancement of conventional and DL approaches for IR, and examination of novel developments in every phase of the processing pipeline. Besides, it presents an analysis of distinct obstacles encountered in iris capture, classification, evaluation, feature encoding, matching, and recognition. Additionally, it evaluates various approaches that show potential for overcoming these challenges. The contributions are determined by the commonly used tools for iris segmentation and recognition using many DL algorithms to improve biometric applications on different sides. It provides an advantage for multi-modal biometrics IR and fingerprint fusion evaluation metrics. Also, it underscores the capabilities of DL techniques in the realm of IR and offers valuable perspectives for forthcoming investigations aimed at enhancing the precision and dependability of biometric systems. The future research directions are identity management and biometric authentication to maximize cyber hygiene. The capacity of on-premise and cloud-based verification solutions to be linked with an organization’s current legacy systems will drive their growing adoption.