**Mohsen Jenadeleh 1,\*, Marius Pedersen <sup>2</sup> and Dietmar Saupe <sup>1</sup>**


Received: 4 January 2020; Accepted: 25 February 2020; Published: 28 February 2020

**Abstract:** Image quality is a key issue affecting the performance of biometric systems. Ensuring the quality of iris images acquired in unconstrained imaging conditions in visible light poses many challenges to iris recognition systems. Poor-quality iris images increase the false rejection rate and decrease the performance of the systems by quality filtering. Methods that can accurately predict iris image quality can improve the efficiency of quality-control protocols in iris recognition systems. We propose a fast blind/no-reference metric for predicting iris image quality. The proposed metric is based on statistical features of the sign and the magnitude of local image intensities. The experiments, conducted with a reference iris recognition system and three datasets of iris images acquired in visible light, showed that the quality of iris images strongly affects the recognition performance and is highly correlated with the iris matching scores. Rejecting poor-quality iris images improved the performance of the iris recognition system. In addition, we analyzed the effect of iris image quality on the accuracy of the iris segmentation module in the iris recognition system.

**Keywords:** biometric recognition; visible light iris images; image quality assessment; image covariates; quality filtering

## **1. Introduction**

The stability of iris patterns over the human lifespan and their uniqueness was first noticed in 1987 [1]. Since then, biometric iris recognition has been extensively investigated for accurate and automatic personal identification and authentication [2]. Most commercial iris recognition systems use near-infrared (NIR) images. However, due to the popularity of smartphones and similar handheld devices with digital cameras, iris recognition systems using images taken in visible light have recently been developed [3–5].

Image quality is a key factor affecting the performance of iris recognition systems [6–8]. In the biometric recognition literature, a biometric quality measure is a covariate that is measurable, influences performance, and is actionable [9–11]. Quality measurement can include subject and image covariates. Subject covariates are attributes of a person, which may be properties of subjects such as eyelid occlusion, glare, iris deformation, or wearing of glasses. Image covariates depend on sensor and acquisition conditions, such as focus, noise, resolution, compression artifacts, and illumination effects. In this work,

we develop a real-time quality measure for image covariates as an actionable quality score, e.g., to decide whether an input iris image sample should be enrolled into a dataset or rejected and a new sample should be captured.

The performance of an iris recognition system in visible light suffers from all of the image quality factors mentioned above. To overcome this problem, some researchers have considered image quality in different ways for iris recognition systems [12–17]. However, these systems fall short in two ways:


In this paper, we propose a general-purpose and fast image quality method that aims to assess the distortion of iris images acquired in unconstrained environments. This method can be used for real-time quality prediction of iris images to rapidly filter image samples with poor quality. Iris images with insufficient quality could lead to high dissimilarity scores for matching pairs and increase the false rejection rate of an iris recognition system. We investigate the effect of iris image quality on the recognition performance of a reference iris recognition system for three challenging iris image datasets acquired in visible light.

This paper is an extended version of our conference paper [18] and mostly a part of the Ph.D. thesis of the first author [19]. The remainder of the paper is organized as follows: Section 2 surveys the literature on iris image quality assessment and iris recognition systems. Section 3 presents the proposed metric for iris image quality assessment. In Section 4, experiments are conducted to study the effect of image quality on the accuracy of iris segmentation. In Section 5, the improvements achieved by filtering poor-quality iris images are discussed using three performance measures on three large iris image datasets acquired in visible light. The paper concludes with suggestions for future research in Section 6.

#### **2. Related Work**

In this section, we review the literature on iris image quality assessment, followed by a brief overview of some state-of-the-art iris recognition systems.

Recently, research has been reported to improve the performance of iris recognition systems by considering image quality, but with certain limitations. In some studies, image quality has been examined by considering only certain quality factors, such as sharpness [20], out-of-focus [21], and JPEG compression [22]. These metrics alone cannot be expected to produce reliable quality assessments of authentic in-the-wild iris images.

In some other work, iris image quality metrics are applied after segmentation of the iris. In [23], the result of the iris segmentation module is used to form a quality score. Happold et al. [24] proposed a method for predicting the iris matching scores of an iris image pair based on their quality features. They calculated these features for precisely segmented iris images. They labeled a dataset of iris image pairs with the corresponding matching scores. They trained their method for predicting the matching score of an image pair based on their quality features. Therefore, these methods cannot be used to measure iris image quality in the iris recognition system pipeline before segmentation.

Several metrics for iris image quality were developed based on a fusion of several quality measures of image and subject covariates. The authors of [25,26] combined quality measures relating to motion blur,

angular deviation, occlusion, and defocus into an overall quality value of an input iris image. These quality metrics were developed for NIR-based images and compared to traditional NIR-controlled iris image acquisition settings. However, images in visible light and under uncontrolled lighting conditions result in notorious differences in the appearance of the acquired images [3]. Therefore, this method may not be used directly to evaluate the quality of iris images in visible light. Li et al. [27] proposed a method for predicting an iris matching score based on iris quality factors such as motion blur, illumination, off-angle, occlusions, and dilation. This method requires segmented irises to compute some of these quality factors (dilation and occlusions).

The authors of [10] used combined subject and image covariates, such as the degree of defocusing, occlusion, reflection, and illumination, to form an overall quality score. They focused on the evaluation of iris images after iris segmentation, which allows the systems to process images of poor and good quality in the acquisition phase. They considered only a few image covariates for quality estimation.

Proença [3] proposed a metric for the quality assessment of iris images taken in visible light. This metric measures six image quality attributes such as focus score, off-angle score, motion score, occlusion score, iris pigmentation level, and pupil dilation. Then, the impact of image quality on feature matching was analyzed. The results showed a significant performance improvement of the iris recognition system by avoiding low-quality images. However, this method requires precisely segmented iris images, and only the motion-blur score is combined with some quality factors related to the subject's covariates.

The authors in [12] proposed an approach that automatically selects the regions of an iris image with the most distinguishably changing patterns between the reference iris image and the distorted version to compute the feature. The measured occlusion and dilation are combined to form a total image quality score to study the correlation between iris image quality and iris recognition accuracy.

In the approach of [28], the image quality is assessed locally, based on a fusion schema at the pixel level using a Gaussian mixture model, which gives a probabilistic measure of the quality of local regions of the iris image. The local quality measure is used to detect the poorly segmented pixels and remove them from the fusion process of a sequence of iris images.

Recently, many image quality methods have been proposed for perceptual quality assessment of natural images [29–35]. Some of these models use statistics of completed local binary patterns (CLBP) as a part of their feature vectors. In [33], joint statistics of local binary patterns (LBP) and CLBP patterns produced quality-aware features, and a regression function was trained to map the feature space to the perceived quality scores. In [32], features based on several local image descriptors such as CLBP, local configuration patterns (LCP), and local phase quantization (LPQ) were extracted, and then a support vector regressor was used to predict the quality scores. These models are trained to predict the perceptual quality of natural images. Liu et al. [36,37] studied some of these methods for filtering low-quality iris images. This study showed inconsistencies for the predicted quality, e.g., removing more low-quality images did not always increase the performance of the iris recognition system. In addition, they removed the low-quality images for each subject separately. Therefore, the filtered images do not have the same range of quality, and there is no global quality-filtering threshold.

In summary, some of the methods for iris quality assessment, such as [25,26], are proposed for NIR images, and only a few types of distortion are considered. Some other quality metrics, like those in [3,23,24], require a segmented iris image to calculate their quality features. They also take limited distortion types into account and are not expected to work well for quality assessment of authentic iris images taken in visible light in arbitrary environmental conditions. Iris recognition systems based on authentic images will broaden the scope of iris recognition systems, and require more research to develop robust metrics for quality assessment of authentically distorted iris images.

Since we used an iris recognition system as a reference system in this paper, in the following, we briefly review some state-of-the-art iris recognition systems.

The fast iris recognition (FIRE) system for images acquired by mobile phones in visible light was proposed by Galdi et al. [38]. It is based on the combination of three classifiers by exploiting iris color and texture information. Raja et al. [39] proposed a recognition system for iris images captured in visible light. This method extracts deep sparse features from image blocks and the whole iris image in different color channels to form the feature vector for an input iris image. Minaee et al. [40] proposed an iris feature extraction method based on textural and scattering transform features. The principal component analysis (PCA) technique is used to reduce the extracted feature dimension.

Recently, OSIRIS version 4.1, an open-source iris detection system, was proposed by Othman et al. [41]. This system follows the classic Daugman method [42] with some improvements in segmentation, normalization, coding, and matching modules. For iris and pupil segmentation, the Viterbi algorithm is used for optimal contour detection. For normalization, a non-circular iris normalization is performed using the coarse contours detected by the Viterbi algorithm. The coding module is based on 2-D Gabor filters, which are calculated in different scales and resolutions. Finally, the matching module calculates the global dissimilarity score between two iris codes using the Hamming distance. We used this system as a reference iris recognition system.

#### **3. Proposed Method**

In this section, we present our fast and general-purpose method for assessing the quality of iris images acquired in visible light.

Earlier works on iris recognition [42,43] employed block-based operations to obtain iris features. Therefore, we can infer that the most distinctive information in the iris pattern comes from the local patterns of an iris image rather than from global features. Local binary patterns (LBP) and their derivatives have been successfully used in many pattern recognition applications, including texture classification [44–46], image retrieval [47,48], object recognition [49,50], action recognition [51,52], and biometric recognition [53–56].

Most of the LBP-based biometric recognition methods use statistical analysis of local patterns for their feature extraction. Wu et al. [29] showed that image distortions could change the statistics of LBPs. They then examined the statistics of the LBPs to suggest an index for evaluating natural image quality. However, this index does not accurately predict image quality for some common image distortions, such as Gaussian blur and impulse noise.

In the proposed differential sign–magnitude statistics index (DSMI), sign and magnitude patterns are first derived. Then, the statistical characteristics of these patterns are analyzed for their sensitivity to iris image distortion. Statistical features of specific coincidence patterns with high sensitivity to image distortion are identified. A weighted nonlinear mapping is applied to the features to form the iris image quality score. This metric takes advantage of the observation that low-quality iris images have fewer of these patterns compared with those in high-quality iris images.

#### *3.1. Proposed Quality Metric*

Our iris image quality metric uses statistical features extracted from patterns of signs and magnitudes of local intensity differences. Then, certain locally weighted statistics of specific sign–magnitude coincidence patterns are used to define the quality score. Guo et al. [46] suggested a completed local binary pattern (CLBP) to represent the local difference information that is missed in the LBP representation of an image [57]. We investigate how common distortions in iris images could alter the statistics of the CLBP. Then, a quality metric based on a specific coincidence of sign and magnitude patterns of the CLBP is proposed.

In CLBP, a local grayscale image patch is represented by its central pixel, and the local differences are given by *dp* = *xp* − *xc*, where *xc* = *I*(*c*) is the gray value of the central pixel of the given patch and *xp* is the gray value of a pixel in the neighborhood. A local difference *dp* can be decomposed into two components, its sign and its magnitude. These signs and magnitudes of local differences are combined into corresponding patterns, CLBP-S and CLBP-M, as follows.

Let *C* = {(*i*, *j*)|*i* = 0, ··· , *M* − 1, *j* = 0, ··· , *N* − 1} be the set of pixels of a normalized grayscale image *I* of *N* pixels width by *M* pixels height. For a given pixel *c* ∈ *C*, let *xc* and *xp*, *p* = 0, ··· , *P* − 1, denote the gray values of the center pixel *c* and the *P* points on a circle of radius *R* about *xc*. For example, suppose the coordinates of *xc* are (0,0); then, the coordinates of *xp* are (*R* cos(2*πp*/*P*), *R* sin(2*πp*/*P*)). The grayscale value *xp* is estimated by interpolation if its coordinates do not coincide with the center of a pixel. Then, the CLBP-S patterns are defined by

$$\text{CLBP-S}\_{P,R}(\mathbf{c}) = \sum\_{p=0}^{P-1} b\_p \cdot \mathbf{2}^p, \quad b\_p = \begin{cases} 1 & \mathbf{x}\_p \ge \mathbf{x}\_{\mathcal{L}} \\ 0 & \text{otherwise} \end{cases}.\tag{1}$$

The CLBP-S operator generates the same code as that of the original LBP operator. The CLBP magnitude patterns are defined similarly by

$$\text{CLBP-M}\_{P,R}(\mathbf{c}) = \sum\_{p=0}^{P-1} b\_p \cdot 2^p, \quad b\_p = \begin{cases} 1 & m\_p \ge z \\ 0 & \text{otherwise}, \end{cases} \tag{2}$$

where *mp* = |*xp* − *xc*| is the magnitude of the local difference *dp*. Furthermore, the threshold value *z* is the average local difference in the *P*-neighborhoods of all center pixels together, i.e.,

$$z = \frac{1}{|\mathbb{C}|P} \sum\_{\mathfrak{c} \in \mathbb{C}} \sum\_{p=0}^{P-1} |\mathbf{x}\_p - \mathbf{x}\_{\mathfrak{c}}|.\tag{3}$$

For each pixel *c* ∈ *C*, we consider the *P*-bit binary representation of the sums in Equations (1) and (2) as binary codes of CLBP-S*P*,*<sup>R</sup>* and CLBP-M*P*,*R*. Using these binary representations, we define rotation invariant indices or patterns for CLBP-S and CLBP-M in a manner similar to that proposed by Ojala et al. [57] for LBP codes. Equation (4) gives the rotation invariant indices of CLBP-S,

$$\text{CLBP-S}\_{P,\mathcal{R}}^{\text{in2}}(\mathbf{c}) = G(\text{CLBP-S}\_{P,\mathcal{R}}(\mathbf{c})) = G\left(\sum\_{p=0}^{P-1} b\_p 2^p\right) = \begin{cases} \sum\_{p=0}^{P-1} b\_p & \text{ll}(\sum\_{p=0}^{P-1} b\_p 2^p) \le 2\\\ P+1 & \text{otherwise} \end{cases}.\tag{4}$$

Here, *U* gives the number of bit changes (0 to 1 or 1 to 0) of the *P*-bit binary representation of a number (including circular shift),

$$M\left(\sum\_{p=0}^{P-1} b\_p \mathbb{2}^p\right) = \sum\_{p=0}^{P-1} |b\_p - b\_{\text{mod } (p+1, P)}| \cdot $$

Similarly, Equation (5) gives the uniform rotation invariant patterns of CLBP-M.

$$\text{CLBP-M}\_{P,R}^{\text{ri}\mu2}(\mathcal{c}) = \mathcal{G}(\text{CLBP-M}\_{P,R}(\mathcal{c})) \,\text{.}\tag{5}$$

Note that these indices, CLBP-S*riu*<sup>2</sup> *<sup>P</sup>*,*<sup>R</sup>* and CLBP-M*riu*<sup>2</sup> *<sup>P</sup>*,*<sup>R</sup>* , range over the set {0, ..., *<sup>P</sup>* <sup>+</sup> <sup>1</sup>}. The first indices from 0 up to *P* correspond to local sign and magnitude patterns with only, at most, two bit changes and, thus, denote uniform local patterns. All non-uniform patterns are assigned to the remaining index *P* + 1.

CLBP-S*riu*<sup>2</sup> *<sup>P</sup>*,*<sup>R</sup>* generates fewer codes than the basic CLBP-S. It carries less textural information by simplifying the local structure. CLBP-M*riu*<sup>2</sup> *<sup>P</sup>*,*<sup>R</sup>* provides a compact representation of textural information derived from local magnitude patterns.

For an illustration for the case of *P* = 4 neighbors at distance *R* = 1 from the central pixel of a patch, we provide Figure 1. We obtain six indices *k* and *l* for sign and magnitude patterns, corresponding to five rotation invariant uniform patterns (*k*, *l* = 0, ..., 4) and one index (*k*, *l* = 5) that represents all non-uniform patterns.

**Figure 1.** The patterns in the upper row correspond to CLBP-S*riu*<sup>2</sup> 4,1 , which compares the gray value of the central pixel of a patch (*xc*) with the gray values of its four neighbors (*xp*). The black and white disks denote smaller and greater values than those of the central pixel value, respectively. In the lower row, CLBP-M*riu*<sup>2</sup> 4,1 compares the absolute values of the differences of the gray values of the central pixel and its neighbors (|*xc* − *xp*|) with the threshold *z* from Equation (3). The hatched and white disks denote smaller and greater absolute values than those of the threshold, respectively. Note that the patterns are rotation invariant. Thus, in the case of *P* = 4 shown here, the patterns for *k*, *l* = 1, 2, 3, 5 may be rotated by multiples of 90 degrees without changing the values of CLBP-S*riu*<sup>2</sup> 4,1 and CLBP-M*riu*<sup>2</sup> 4,1 .

Finally, the local indices for sign and magnitude have to be combined to give a quality indicator for an iris image as a whole. We first join the two types of indices into a set of bitmaps *Vk*,*l*(*c*), indexed by *k*, *l*,

$$V\_{k,l}(\mathbf{c}) = \begin{cases} 1 & \text{CLBP-S}^{\text{ru2}}\_{P,\mathcal{R}}(\mathbf{c}) = k \text{ and } \text{CLBP-M}^{\text{ru2}}\_{P,\mathcal{R}}(\mathbf{c}) = l \\ 0 & \text{otherwise} \end{cases} \tag{6}$$

For each pair *k*, *l* of indices, we form a weighted sum of *Vk*,*l*(*c*) over all pixels *c*, which is nonlinearly scaled to the unit interval by *<sup>r</sup>*(*x*) = <sup>1</sup> − *<sup>e</sup>*−*ax* as follows:

$$Q\_{k,l} = r \left( \frac{1}{|\mathbb{C}|} \sum\_{\boldsymbol{c} \in \mathbb{C}} \frac{V\_{k,l}(\boldsymbol{c})}{\hat{\sigma}^2(\boldsymbol{c}) + \delta^2} \right). \tag{7}$$

Here, *σ*ˆ <sup>2</sup>(*c*) is the local variance of the *P*-neighboring pixels of the center pixel *c*, and *δ*<sup>2</sup> is a small constant value to prevent division by zero. The parameters *δ*<sup>2</sup> and *a* are empirically set to 0.00025 and 0.01, respectively.

In Equation (7), the normalization by the local variance emphasizes local minima and maxima, and normalizing the scores to the range [0, 1) is only for ease of interpretation of the quality scores. The value of *Qk*,*<sup>l</sup>* is considered as an image quality score derived from the sign pattern *k* and the magnitude pattern *l*. In our experiments, we used four neighbors (*P* = 4) with unit distance (*R* = 1) from the central pixel *c* of a local patch.

Our experiments showed that *Qk*,*<sup>l</sup>* with the specific coincidence of the sign pattern *k* = 0 and magnitude pattern *l* = 0 has a high correlation with iris image quality. Therefore, we used *Q*0,0 as our proposed DSMI quality score. We had summarized the proposed DSMI metric in our conference paper [18], considering, however, only the selected coincidence sign–magnitude patterns.

#### *3.2. Empirical Justification*

Inspired by Wu et al. [29], we examine the distinctiveness of each pattern of CLBP-S*riu*<sup>2</sup> 4,1 , which coincides with patterns of CLBP-M*riu*<sup>2</sup> 4,1 for separating high-quality iris images from distorted versions. To that end, we generated an artificially distorted iris image dataset from 600 pristine high-quality references taken from the Warsaw-BioBase-Smartphone-Iris v1.0 [4], UTIRIS [58], and *GC*<sup>2</sup> multi-modal [36] datasets. A total of 3 to 12 samples per eye from 75 individuals were selected. This dataset was used only to justify our choice of specific sign–magnitude patterns and also to investigate how filtering out the low-quality iris images using the DSMI metric could affect the performance of the segmentation module of the reference iris recognition system. The reference iris images have no content-dependent deformations such as eyelid occlusion, and were selected from individuals with high, medium, and low degrees of iris pigmentation. The irises of all of these reference iris images were segmented accurately by the reference iris recognition system.

Five common image distortions with different levels and multiple distortions were used to distort the reference iris images. These distortions are Gaussian blur (GB), motion blur (MB), white Gaussian noise (WGN), salt and pepper noise (IN), and overexposure (OE). The parameters of each function and the number of the distorted versions of each reference image are listed in Table 1. In addition to the individual types of distortions, we generated multiple distorted iris images (GB+WGN). First, we distorted the images with GB and then with WGN. Since GB tends to occur during the acquisition phase due to the different working conditions of the image sensors, we applied it first. WGN is a noise model that can be used to mimic the effects of random processes, such as sensor noise due to poor illumination and thermal noise in the imaging device. For simplicity, the recommendation of [59] was followed, and WGN was introduced in the end.


**Table 1.** Summary of the artificially distorted iris image dataset.

To analyze the discrimination power of the scores *Qk*,*<sup>l</sup>* for separating the high-quality reference images from their distorted versions, we show the distributions of the corresponding scores *Qk*,*<sup>l</sup>* for some selected combinations of *k* and *l* in Figure 2. Visual inspection clearly shows that the coincidence of sign–magnitude patterns with *k* = 0 and *l* = 0 gives the greatest discrimination power. The predicted quality scores for the

reference iris images are mostly between 0.8 and 1, and the scores for the distorted versions are mostly less than 0.8. Therefore, we chose this coincidence pattern to form our DSMI quality metric (DSMI = *Q*0,0).

**Figure 2.** The solid red lines show the distributions of the quality scores of the high-quality iris images, and the dotted blue lines show the distributions for the distorted versions with different distortion types, which are shown on the right side of each row. The quality scores *Qk*,*<sup>l</sup>* are formed based on four different coincidences of sign (*k*) and magnitude (*l*) patterns, shown at the bottom of each column. The first column shows the histograms of the quality score *Q*0,0, and the second, third, and fourth columns show the histograms of the coincidence patterns *Q*0,*<sup>l</sup>* with *l* = 0, *l* = all, and *l* = 4.

#### **4. Iris Segmentation Accuracy**

The performance of iris segmentation in a classical iris recognition system has a significant impact on the overall performance. In this section, we analyze how image distortions affect the performance of the segmentation module and how quality filtering could improve the segmentation.

Most of the state-of-the-art iris recognition systems for iris imaging acquired in visible light, such as FIRE [38], Raja et al. [39], and OSIRIS, version 4.1 [41], can be used as reference iris recognition systems. We have chosen OSIRIS version 4.1 because (1) OSIRIS is an open-source iris recognition system that facilitates reproducible experiments, (2) it shows high recognition performance [41], and (3) it was used as the reference iris recognition system in some recent biometric recognition studies [4,60–64]. The segmentation module of OSIRIS version 4.1 uses the Viterbi algorithm to detect the iris and pupil contours [65]. The outputs are contours of the iris, which represent the inner boundary between the pupil and iris and the outer boundary between the iris and sclera, resulting in a binary mask for the iris.

For our experiments, we used the artificially distorted dataset from the previous section, which is summarized in Table 1. We segmented all iris images using the OSIRIS segmentation module. The mask of the segmented iris of each reference image was taken as the ground truth for comparison with the segmentation results for the distorted versions. The iris segmentation error is computed by the fraction of mislabeled pixels,

$$\mathcal{c} = \frac{1}{|\mathbb{C}|} \sum\_{\mathfrak{c} \in \mathcal{C}} T(\mathfrak{c}) \oplus M(\mathfrak{c}),$$

where |*C*| is the cardinality of the pixel set *C* of an iris image, and *T* and *M* represent the ground truth and the generated iris masks, respectively. The symbol ⊕ represents the exclusive OR operation to identify the segmentation error. If the error *e* was below the threshold 0.05, the iris segmentation was assumed to be correct. The threshold value was set manually by the authors.

In Figure 3, we show the fractions of incorrectly segmented irises for the different types of distortion and for low, medium, and high degrees of iris pigmentation. The fractions are given as functions of the percentage of low-quality images that were filtered out using the proposed DSMI quality metric.

The results shown indicate a clear correlation between the DSMI quality of iris images and segmentation accuracy. Therefore, filtering out poor-quality images before segmentation will improve the performance by reducing the number of incorrectly segmented images, as indicated by the negative slopes of the plots.

In summary, the experiments performed in this section show that the accuracy of the segmentation module varies for iris images with different pigmentations and different distortions. Highly pigmented iris images present a greater challenge for the reference iris recognition system, while the system is more robust for the segmentation of low-pigmented iris images. However, filtering out poor-quality iris images using the proposed DSMI metric increases the accuracy of iris segmentation.

**Figure 3.** The segmentation performance of the reference iris recognition system is shown for segmenting iris images with high, medium, and low pigmentation, and distorted in different ways. The fraction of incorrectly segmented images is plotted versus the percentage of filtered low-quality images, based on the differential sign–magnitude statistics index (DSMI) metric.

#### **5. Experimental Results**

In this section, we investigate to what extent filtering out poor-quality iris images with the proposed quality metric improves the performance of the reference iris recognition system. We also compare our DSMI quality metric with the BRISQUE [66] and WAV1 [67] image quality metrics. BRISQUE uses statistical features extracted from pixel intensities to train a support vector machine for predicting image quality. Pertuz et al. [67] compared 15 metrics to estimate the blur of an image. In their study, WAV1 performed better than the others. WAV1 uses statistical properties of the discrete wavelet transform coefficients. Since blur is a common distortion of iris images taken by handheld imaging devices such as smartphones, we also compare our method with the WAV1 metric. Our experiments were conducted on three large authentic iris image datasets acquired in visible light.

#### *5.1. Iris Image Datasets*

There are many iris image datasets recorded with near-infrared cameras such as CASIA V4 [68], CASIA-Iris-Mobile-V1 [69], IIT Delhi [70], and ND CrossSensor Iris 2013 [71]. However, there are just a few iris image datasets acquired in visible light. Four are widely used in iris recognition research: UTIRIS [58], UBIRIS [72], MICHE [73], and VISOB [74].

An optometric framework in a controlled environment was used for capturing the irises of the UTIRIS dataset, resulting in high-quality iris images. UBIRIS iris images were taken from moving subjects and at different distances, resulting in more heterogeneous images compared to UTIRIS. Nevertheless, the pictures have good quality, better than the expected quality of iris images captured by handheld devices. The MICHE and VISOB datasets are challenging datasets for iris recognition systems, including images with varying degrees of iris pigmentation and eye make-up. In addition, the quality of the images is

impaired by lack of focus, gaze deviations, specular reflections, eye occlusion, different lighting conditions, and motion blur.

Instead, we chose three datasets of the *GC*<sup>2</sup> multi-modal biometric dataset [36] because they contain authentically distorted iris images typically seen when capturing iris images with handheld devices such as smartphones. In addition, the iris images were taken from many subjects with different handheld cameras in uncontrolled environments at different distances. Iris pigmentation varied, from European subjects with bright iris textures to Asian subjects with very dark iris textures. In addition to the various authentic distortions corresponding to the image covariates, the iris images are subject to a variety of quality losses related to the subject's covariates, such as gaze deviation, off-angle, reflections, eye closure, and make-up. Also, the datasets contain 12–15 iris images of varying quality per eye and person, which is useful for studying the effect of quality filtering. The iris images have more than 30 different resolutions.


We compare an iris image with all iris images from the same dataset. Table 2 summarizes these datasets and shows the number of matching and non-matching iris pairs. Figure 4 shows some samples from these datasets, and Figure 5 shows the histograms of the quality scores of the datasets, estimated by the proposed DSMI metric.

**Figure 4.** Some iris image samples with high, medium, and low pigmentation from the multi-modal biometric dataset *GC*<sup>2</sup> [36]. The first, second, and third rows show some images from the REFLEX, LFC, and PHONE datasets, respectively.


**Table 2.** Summary of the *GC*<sup>2</sup> dataset.

**Figure 5.** Normalized histograms of the quality scores according to the DSMI metric on three test iris datasets.

(**b**) LFC

(**c**) PHONE

#### *5.2. Iris Recognition Performance Analysis*

(**a**) REFLEX

To evaluate the performance improvement of iris recognition achieved by quality filtering using an image quality metric, we used three performance methods, namely the Daugman's decidability index [75], the area under the receiver operating characteristic curves (AUC), and the equal error rates (EER). We compared the performance of three image quality metrics when used for quality filtering. Given a threshold for a metric, we rejected those images that exhibited a quality lower than the threshold. The thresholds for each of the three metrics were chosen such that 1/4, 1/2, and 3/4 of the images were rejected. In our experiments, OSIRIS version 4.1 was used as a reference iris recognition system.
