Implementation of selected ISO/IEC 29794-5 measures and proposing alternatives

1 Introduction

Face recognition is now extremely used across various sectors, helping both to increase the convenience of everyday activities, such as unlocking your smartphone, and to improve security, for example, at the airports. The image quality depends on the conditions under which the picture is taken and plays a crucial role in the effectiveness of the biometric systems. Low image quality directly lowers the security and reliability of the solution. The need for systems prepared for various type of face occlusions became evident during COVID-19 pandemic. Once the mandatory facemasks were introduced, most facial recognition systems either stopped functioning or gave wrong results. This demonstrates a strong demand for such solution and reveals that the existing systems were ill-prepared for potential occlusion cases. One could argue that this was only a temporary issue, connected to the global pandemic. However, it is important to remember that the subtle occlusions can be much harder to detect. Heavy makeup, facial hair or lightning conditions may not seem like an important factor, yet they affect the recognition results more than we can expect. To define the optimal conditions for capturing a biometric sample, the ISO/IEC 29794-5:2023 (International Organization for Standardization and International Electrotechnical Commission, 2023) standard was developed by biometric experts in order to establish guidelines that have to be followed when collecting biometric data. At the time of writing this analysis, work on the standard was still in progress, though a substantial amount of work has been published in the draft document ISO/IEC 29794-5:2023 (CD3) (International Organization for Standardization and International Electrotechnical Commission, 2023), which serves as the basis for all references to the standard ¹ .

In response to critical need for a reliable facial image quality assessment, this work presents a detailed analysis of the ISO/IEC 29794-5:2023 (CD3) standard and the implementation of a software tool for evaluating facial image quality based on selected standard measures. Additionally, alternative program will be developed–consisting of the same set of measures but utilizing state of the art solutions. Both will provide an overall percentage score, consisting of weighted results for individual measures. Based on these scores and own observations, both the solutions will be compared using a set of tests.

The work is divided into six main parts. The previous work chapter (Section 2) focuses on biometric verification and the challenges associated with it. It also examines the current state of the art technologies and discusses image quality concerns in terms of recognition systems.

Methodology chapter (Section 3) describes our approach to data selection and preparation, presents overview of the experimental settings and outlines the evaluation methodology used to both assess the image quality measures and their impact on face recognition performance.

The implementation chapter (Section 4) describes the implementation of selected quality measures compliant with ISO/IEC 29794-5:2023. It outlines the recommendations for each selected measure and explains how the outputs are converted into measurable quality indicators. In addition, an alternative implementation of the standard’s measures is introduced. This chapter further reviews state-of-the-art technical solutions and approaches, as well as conducts a series of performance tests in order to identify the most effective implementation method. A detailed explanation of each method and its conversion to quality score is also provided. The experiments, optimization and results (Section 5) chapter presents the experiments conducted as part of this work. It details the testing methodology, dataset used and test-bench structure. Each quality component is compared, and its impact on the final score is analyzed. Furthermore, thresholds defining three quality levels are verified in order to select the most effective ones. We present and discuss there the results of our experimental settings.

The summary chapter (Section 6) concludes the work, presenting summary of the research and highlighting key takeaways, and proposes directions for further development. Additionally, it outlines the potential industrial and commercial applications of the proposed solution, and the study limitations.

2 Previous work

Face Image Quality Assessment (FIQA) represents a technically demanding niche within the biometrics field, concerned with the question whether the image is suitable for biometric system. Unlike general IQA, FIQA has to take into consideration subtle factors like pose or face occlusions which can significantly reduce recognition performance even if the system seems to be highly effective. Early studies focused on handcrafted features such as sharpness, contrast or noise, which moderately effective, struggled to generalize across unexpected real-world conditions. Khodabakhsh et al. (Khodabakhsh et al., 2019) demonstrated that subjective quality assessment, reflecting human opinions and views, proved to be better indicator than standard metrics, emphasizing the difficulty of assessing the recognition-relevant quality.

The field advanced with deep learning, using models to learn discriminative features from the prepared data. Convolutional Neural Networks became particularly effective, showing good performance across many difficult conditions. For example, advanced frameworks like CR-FIQA (Boutros et al., 2023) predict quality by evaluating a sample in feature space, capturing subtle inter- and intra-class variations. Training-free methods like GraFIQs (Kolf et al., 2024) assess quality through gradient-based analysis to estimate quality without supervision and pre-labeled datasets.

Studies in much larger scale are important in terms of perspective on the topic. The NIST FRVT quality assessment framework ² , including the SIDD report (National Institute of Standards and Technology, 2023), analyzed many image quality algorithms and how they impact face recognition accuracy. The conclusion was that filtering out low quality images improves performance, but most of the evaluation was done on proprietary models, which are not available to the public. That limits reproducibility and prevents researchers from using the best solutions.

At a more general level, the ISO/IEC 29794-1 (International Or ganization for Standardization and International Electrotechnical Commission, 2024) standard on biometric sample quality defines general framework for biometric data quality assessment for different modalities. Character, fidelity and utility are used to distinguish human perceived and system assessed quality. Instead of defining specific metrics, the norm focuses on establishing general principles and terminology, serving as a basis for modality-specific standards.

Combining research and the industry needs, the latest image quality research focuses on standardization and practical deployment. ISO/IEC 29794-5 (International Organization for Standardization, 2025) provides guidelines on measuring face image quality in terms of biometric system usability. Following these standards, the Open Face Image Quality (OFIQ) framework (BSI-OFIQ Project Contributors, 2024a) was developed as an open-source solution. The OFIQ development report (BSI-OFIQ Project Contributors, 2024b) describes how the system works, which quality measures it uses and how it was tested, making the FIQA research more transparent than previous solutions.

Despite these advances, several challenges remain, yet to be solved. The lack of comprehensive datasets stops us from applying the model solutions to be used in real-world scenarios. Moreover, the field distinguishes between two types of face image quality: the human-perceived and the system-assessed, which frequently do not align. Our method aims to showcase this difference.

3 Methodology

The proposed methodology consists of several steps. First, a dedicated face image dataset was constructed and manually annotated into three quality subsets (by human perception of image quality), similar to the approach proposed by Zhang et al. (Zhang, 2023). The subsets represent three levels of face image quality (bad, good, very good) and are the baseline for the quality assessment algorithms. The whole dataset consists of 300 images, and subsets, respectively 100 (for bad), 100 (for good) and 100 (for very good). The face images were taken from two public datasets: Lagenda (Kuprashevich and Tolstykh, 2023) and LFW (Labeled Faces in the Wild) (Huang et al., 2007), samples of the images are presented in Figure 1.

In the second phase, we pick the five selected quality measures from ISO/IEC 29794-5:2023 - single face present, natural colour, eyes open, pose and moments of luminance distribution. The selection was made based on the frequency of occurrence of a given flaws in selfies used for identity confirmation in thousands eKYC (electronic Know Your Customer) attempts in high risk usage scenarios (finance). We gain access to these data thanks to the close collaboration with an industry partner. We inspected the recorded data manually and observed these property.

Next, two comparable face image quality systems were implemented considering the aforementioned five selected quality measures from ISO/IEC 29794-5:2023. The first system is created according to the standard’s recommendations, while the other is based on our solutions.

In order to get the final quality score we evaluate each measure based on the ability to classify the image into the correct quality subset. This evaluation is performed independently for each measure and decision thresholds are determined to minimize misclassifications errors among the quality levels. Based on the performance, the measures are weighted and combined to produce a final quality score.

Face recognition systems are known to be sensitive to low-quality input images and such examples often lead to incorrect decisions, increasing the False Non-Match Rate (FNMR) and False Match Rate (FMR). Based on this observation, the impact of rejecting low quality images on recognition performance is evaluated.

The another evaluation dataset is prepared based on 2000 images, which consists of 200 subjects and 10 samples in the single class. We use the VGGFace2 dataset (VGGFace2 Dataset, 2018), which contains face images of 480 people compiled from internet sources, samples of the images are presented in Figure 2. The images vary in terms of race, pose, illumination, and other quality factors discussed earlier (see Section 4). The dataset was originally created and curated using a data-driven approach to clean large face collections (Cao et al., 2018), ensuring diversity images suitable for face recognition research. We pick off-the-shelf face recognition algorithm - AdaFace (Kim et al., 2022) considering its performance and accessibility.

First, reference recognition performance is measured without discarding any samples. We apply a face recognition system on the evaluation dataset and derive fixed decision threshold from the ROC curve (at EER = 0,06). This point is used as a reference for all further experiments. Next, images with quality score below a given threshold are gradually removed from the dataset and face recognition process is conducted. We increase the image quality threshold by given small step - 5, while computing the FNMR and FMR for the remaining images, following each step, simultaneously recording the discarded image percentage. Due to the effect that a fraction of low-quality samples are excluded in every iteration step, the FNMR should decrease constantly if the quality measure is a good predictor for the biometric performance (Priesnitz et al., 2024).

This evaluation approach closely relates to the method of Error vs. Reject Curve (ERC) (Grothe et al., 2007) and the idea of Error vs. Discard Characteristic (EDC) curves, commonly used in biometric quality assessment. As described in a paper (Schlett et al., 2023) and mentioned in ISO/IEC 29794-1:2024(8), EDC curves analyze recognition error rates and the amount of discarded samples. Consequently, the proposed methodology is consistent with the evaluation practices and allows for systematic assessment of the image quality measures impact on the face recognition performance.

4 Implementation

The analyzed standard proposes a range of facial image quality measures, which are highly diverse both in the evaluated features and implementation methods. From these, five measures were selected that, in authors’ opinion, provide protection against the widest range of biometric system errors (see 3). The selected measures include: single-face presence, eye openness, natural facial color, face orientation, and luminance distribution. Some of the chosen methods required additional functionalities, such as landmark detection or face segmentation. These supporting functions were integrated as an essential part of the quality assessment system.

The measure implementation can be divided into two separate stages, differing in the approach used during the creation process. The standard-compliant version required the conversion of descriptive content and mathematical formulas into executable code. The custom implementation involved a comprehensive analysis on the latest technical solutions, followed by testing and selecting the most appropriate methods for each case.

Each measure was designed to return normalized quality score ranging from 0 to 100, where the values correspond to very poor and perfect conditions for biometric system. To create a coherent overall evaluation, the results of individual measures were aggregated using a weighted mean, producing the final quality score. The following formula was applied in Equation 1:

Q C final = Q C 1 ⋅ w 1 + Q C 2 ⋅ w 2 + ⋯ + Q C 5 ⋅ w 5 (1)

Where the variables Q C (quality check) represent the individual measure result, and the w represents the corresponding weight. For easier score interpretation, the weights were normalized so that the final result also remains within the 0-100 range. The weighted overall score was then translated into descriptive score, easily interpreted by users, according to the logic defined in Equation 2:

Q C descriptive = bad , for Q C final < a good , for a ≤ Q C final ≤ b very good , for b < Q C final (2)

Where parameters a and b define the thresholds for transitions between the descriptive quality categories. This approach leaves two factors requiring experimental verification: the weights of individual measures and the thresholds of the final scale. Consequently, a comprehensive testing procedure was developed to evaluate the influence of individual measures on quality assessment accuracy and to analyze the performance of different final evaluation thresholds. Below we showcase a summarized implementation of each set of the measures, both standard-compliant and our implementation.

4.1 Single face present

The standard specifies that only one face should be visible in the image; however additional and undesired people can appear in the background - either as bystanders, images or prints on clothing. The system may detect such faces and mistakenly use them for further processing, inevitably resulting in an error.

4.1.1 Standard compliant implementation

The process begins with implementing the face detector recommended by the standard—SSD (Single Shot Detector) based on a convolutional neural network (Rath, 2024). The detector’s parameters are consistent with original implementation; however, it should be noted that the model’s confidence threshold is supposed to be set at 50% which increases the number of false detections. After the image has been processed, the output consists of bounding box coordinates, corresponding to detected faces. Depending on the number of detected people, the parameter f - face unicity is calculated according to Equation 3:The parameter T denotes the pixel distance from the chin to the midpoint of the eyes. The value is computed for each detected face and the largest value T 1 and second-largest T 2 are selected to continue the computation. The final step is calculating the Q C - quality score using the following formula shown in Equation 4:

f = 1 for no faces detected 0 for single face detected T 2 T 1 for multiple faces detected (3)

Q C = ROUND 100 1 − f (4)

4.1.2 Custom implementation

Given the vast diversity of existing algorithms, only two implementations were selected for comparison: the OpenCV library (OpenCV, 2000) and the MediaPipe framework (Google, 2019). The choice was partly influenced by their popularity and community support, which helps with practical integration and troubleshooting. The primary motivation was the technical contrast between these two methods. While MediaPipe relies on deep neural networks, OpenCV uses classical algorithms like Haar Cascades (Viola and Jones, 2001). This allows us for a direct comparison between traditional and modern approaches to face detection.

4.1.2.1 OpenCV

OpenCV is open source library for computer vision that can be used for various tasks like face recognition, motion tracking or image segmentation. It uses classical algorithms rather than deep learning which allows for efficient execution, without high computational requirements. To use the face detector we start with loading the library, then preprocessing the image, loading a model (e.g., Haar Cascades) and applying detection to obtain bounding box coordinates.

4.1.2.2 MediaPipe

MediaPipe is a framework developed by Google., utilizing deep neural networks for image analysis. It enables quick implementation of advanced AI models, including motion and gesture analytics. Optimized for devices with limited computational power, face detection involves loading the library, optionally processing the image, configuring the model and retrieving bounding box coordinates. Additionally, MediaPipe offers a highly accurate facial landmark detection function, generating precise face mesh.

4.1.2.3 Method selection

When comparing both methods we focused on several key factors: ease of implementation, execution speed, detection quality and additional features. Implementing both tools was relatively easy, following advice and using the community support. Tests on images of various quality showed no significant performance differences between the two detectors, once the detection thresholds were appropriately adjusted. The only difference was execution time: MediaPipe ran slightly slower due to more complex model. The most important aspect were the additional features. High-quality landmark mapping was necessary. For later processing stages and while OpenCV offers 68 facial landmarks, MediaPipe provides 468, offering a much higher detail. Consequently, that was the final tool which we implemented.

The method was implemented following the previous description. Then, depending on the amount of detected faces n, the quality score was computed according to Equation 5:

Q C = 0 for no faces detected 100 for single face detected 100 n for multiple faces detected (5)

4.2 Natural colour

The natural skin colour is very sensitive topic due to diversity of skin tones across population. It is essential to ensure that the component will not lead to bias or unjustified restrictions. To prevent this, the authors of the standard extended the range of acceptable values compared to other studies. Using these wider limits allows for better reliability, both in terms of natural skin color variations and additional factors like lighting conditions.

4.2.1 Standard compliant implementation

The computation starts by finding the parameter D representing the degree of skin tone unnaturalness, using Equation 6:

D = 0 for single luminance channel and no chrominance channels 0 when all pixels ′ R = G = B (6)

In other cases, R , G , B values are computed for each pixel and are transformed into CIELAB colour space. From this colour space we will use the a * and b * components. If both values are below zero D = 100. Otherwise, it is calculated as the distance in the ( a * , b * ) space from the skin tone region, according to Equation 7:

D = max max 0,5 − a * , max 0 , a * − 25 2 + m a x max 0,5 − b * , max 0 , b * − 35 2 (7)

Using the calculated D, the final quality score can be computed as shown in Equation 8:

C = ROUND 100 1 − SIGMOID D , 0,10 (8)

4.2.2 Custom implementation

Although seemingly niche, natural skin colour detection is used in identity verification, face tracking and even medical and cosmetic applications. It is a very subtle therefore allows for little margin for error under difficult conditions such as poor lighting or noisy images. In this case we compared two methods: colour space analysis and deep learning. Three colour spaces - RGB, HSV, and YCbCr were analyzed for accuracy under varying conditions. Deep learning was included as an alternative, often used in medical and dermatological context.

4.2.2.1 RGB, HSV and YCbCr colour spaces analysis

RGB is the most commonly used colour space, representing colour through three channels (Red, Green, Blue). Implementation involved face region detection and converting it to RGB using OpenCV. Pixel values within the predefined skin colour range (Bencheriet et al., 2010) were counted and compared to the total number of pixels. Regions with more than 70% match were considered natural. While simple and intuitive, it proved to be less effective for accurate skin colour, especially under random colour and changing lighting conditions.

HSV separates brightness from hue, allowing for minimization of lighting impact. After transforming the detected face area to HSV space, pixel ratios within the predefined skin tone range (Cheref, 2024) were calculated in the same manner. Tests confirmed high precision and reliability in detecting skin, even in difficult conditions and unnatural shades.

YCbCr also separates brightness from chrominance. Following the same procedure - face extraction, conversion, and pixel ratio analysis. It performed well in tests, though some blocky artifacts appeared, which could cause errors in cases of partial occlusion (e.g., glasses or masks).

4.2.2.2 Deep learning model

Deep learning models can be highly effective, capable of identifying subtle skin variations and are often applied in dermatological diagnosis (Santos et al., 2020). They improve the detection of fine details, often overlooked by human eye. To implement such model, a large dataset with skin segmentation (e.g., facial or hand regions) must be gathered. The model is trained by passing these images through neural network, adjusting its weights to better identify the target regions. The process is complex and demands a large amount of computational and time resources.

4.2.2.3 Method selection

Since the deep learning models are usually applied for fine details rather than general colour verification, the method was deemed inefficient for our purpose due to its resources dependency. Considering the analyzed colour spaces, HSV was chosen as the best option. It captured the most detail, was resilient to varying light and unnatural tones, and required minimal computational resources.

The implementation follows the process described earlier. After obtaining the proportion of skin-coloured pixels in the selected region - P, the quality measure can be obtained as shown in Equation 9:

Q C = P 5 f o r 0 % < P ≤ 70 % 100 f o r P > 70 % (9)

4.3 Eyes open

Eyes are one of the most distinctive features of the human face, making them crucial for biometric identification systems. Therefore, any deviation from their natural state such as complete closure or abnormally narrow or wide opening, may lead to errors mainly from the absence of the iris and displacement of distinctive facial features.

4.3.1 Standard compliant implementation

To accurately identify points around the eyes, the process starts with facial detection and then determines facial landmarks via the ADNet (Huang et al., 2021) neural network model. Based on the detected landmarks, three crucial distances are calculated: T, representing the number of pixels measured from the chin to the point between eyes; D L, the maximum distance between the upper and lower eyelids of the left eye; and D R the maximum distance between the upper and lower eyelids of the right eye. Next, the smaller value between D L and D R is selected and labeled as D PAL. This enables the calculation of eye openness using the following Equation 10:The obtained result is then applied in the Equation 11 used to determine the quality measure:

ω = D PAL T (10)

Q C = ROUND 100 1 − SIGMOID ω , 0.2 , 0.06 (11)

4.3.2 Custom implementation

For the comparison, three algorithms were selected, each representing different approach to the problem: facial landmark detection, iris detection and color proportion evaluation. Each focuses on a different feature, allowing for variable performance under different illumination and occlusion conditions.

4.3.2.1 Computation from facial landmarks

Mapping facial landmarks allows for accurate identification of facial key points, including those difficult to obtain, such as the upper and lower eyelids edges. Therefore, the MediaPipe library was used to create a mesh of mapped facial landmarks. After identifying points around the eyes, the ones along the upper and lower eyelids were extracted. The inter-eyelid distance was then measured and normalized by the height of the detected face to obtain the eye-to-face proportion denoted as D. The value was computed following the Equation 12:

D = x 2 − x 1 2 + y 2 − y 1 2 h (12)

where ( x 1 , y 1 ) and ( x 2 , y 2 ) are the coordinates of the lower and upper eyelid points, while h stands for face height. This computation was performed for both eyes. Based on the results, we determined the percentage of the face area covered by open eyes, defining the verification threshold. After testing on a diverse dataset of facial images varying in lightning, person’s age and cultural background, the following thresholds were established using Equation 13:

D = D ≤ 0.05 eyes closed 0.05 < D ≤ 0.15 eyes open D > 0.15 false detection (13)

4.3.2.2 Color proportion analysis

The usage of color analysis allows for distinguishing between open and closed eyes efficiently, even in cases of anatomical variations such as diverse eyelid shapes. The process focuses on the sclera color, since white does not appear anywhere else on the face in natural setting. Using previously detected landmarks, the eye region was defined and converted to HSV color space, which allowed for effective separation of light tones from the rest of the region. From the extracted region, both the total number of pixels ( n region ) and white pixel count ( n white ) were obtained in order to compute their ratio (P) calculated as shown in Equation 14:

P = n white n region (14)

The computed value was then compared with predefined thresholds as shown in Equation 15. The thresholds were established based on author’s own analysis and a series of tests conducted on a diverse dataset of facial images, varying in conditions that could influence the results.

Q C = P < 0.3 eyes closed 0.3 ≤ P ≤ 0.6 eyes open P > 0.6 false detection (15)

4.3.2.3 Iris detection

Although iris position analysis was not originally designed for eye openness detection, it serves as a practical alternative. The method focuses on the iris shape recognition and shows resistance to different lightning and anatomical conditions. The implementation reuses the MediaPipe library, utilizing the detected face and its bounding box, focusing only on the upper part since that where the eyes are located. The area is converted to grayscale, noise is reduced via Gaussian filtering, and iris circles are identified thanks to properly configured Hough Transform (Biranje et al., 2019).

4.3.2.4 Method selection

Analysis of the three approaches showed that each of them performs better under different conditions due to their unique characteristics. Landmark-based calculations were reliable in standard conditions, but unusual eye shapes caused significant difficulties. Ther proportion method performed best in uniform lightning, yet struggled in overexposure and underexposure cases. Iris position analysis stood out the most, accurately detecting natural eye shapes even in cases challenging for the other methods. Although it does not directly verify the eye openness, it was most effective for estimating the required verification parameter.

The method was implemented following the previous description. After obtaining the number of detected circles n, the quality measure was computed as shown in Equation 16:

Q C = 0 f o r n = 0 80 f o r n = 1 100 f o r n = 2 100 n f o r n > 2 (16)

4.4 Pose

A non-frontal orientation of the face relative to the camera significantly affects the facial recognition process. This occurs both due to partial occlusion of key facial features and because reference images are always captured with the face directly facing the camera. Head orientation can be defined by three angles - yaw, pitch and roll - corresponding to rotations around the Y, X and Z-axes respectively.

4.4.1 Standard compliant implementation

To define the degree of head tilt, the SynergyNet (Wu et al., 2021) neural network model was used. Based on the model’s output, the head rotation matrix was computed, allowing for extraction of the angles relative to optical axis - ϕ yaw , ϕ pitch , ϕ roll. These allow for computing the corresponding quality verification metrics.

4.4.1.1 Yaw

Yaw represents head rotation around the Y-axis, indicating whether it is turned left or right. Large deviations obscure key facial features, making the recognition very difficult. The corresponding quality score for Y-axis rotation is computed as shown in Equation 17:

Q C = ROUND 100 max 0 , cos ϕ Y AW (17)

4.4.1.2 Pitch

Pitch represents the tilt of the head along the X-axis, indicating whether it is directed down or upward. Unlike yaw, defining “neutral” position in this case is much more difficult. Nevertheless, for correct recognition it is crucial for correct face recognition to keep the pitch values small, ensuring alignment with reference picture. The quality score is computed as shown in Equation 18:

Q C = ROUND 100 max 0 , cos ϕ PITCH (18)

4.4.1.3 Roll

The final component describes the head rotation in the Z-axis, corresponding to lateral or diagonal head tilt. A correct angle ensures that the facial features are symmetrically aligned, improving the recognition process. The quality score for Z-axis rotation can be determined using Equation 19:taking into consideration that for | ϕ ROLL | ≥ 90 the Q C = 0.

Q C = ROUND 100 max 0 , cos ϕ ROLL (19)

4.4.2 Custom implementation

We decided to compare three approaches: two-dimensional landmark analysis, three-dimensional landmark mapping and pretrained neural network. The 2D approach uses only key facial points for fast processing. The 3D facial points mapping allows for more detailed approach, taking into consideration perspective and additional dimension. In contrast, the neural network uses advanced machine learning models to analyze the face under diverse conditions.

4.4.2.1 Facial landmarks position analysis

The simplest method for frontal face detection is 2D analysis of facial landmarks. Assuming precise mapping, deviations along X, Y and Z-axes are estimated by comparing landmarks laying on the edges of the face. The approach is pretty straightforward but can be prone to errors due to lack of depth information. Implementation involves MediaPipe-detected key points: the nose tip, the outer eye corners, and the midpoint between the eyes. Rotation angles can be derived from these points: yaw from the nose-eye line, pitch from vertical differences, and roll from eye corner alignment. Tests showed that while the method is rapid, it is unreliable, highly sensitive to lighting, environmental conditions or individual facial structure with asymmetries strongly affecting performance.

4.4.2.2 3D mapping of facial landmarks

While 2D mapping does not directly represent 3D face position, mathematical transformations allow projection into 3D space, enabling detection of deviations from the frontal position. Implementation begins by defining a standard 3D face (LearnOpenCV, 2024), using points such as the nose tip, chin and eye corners. Landmarks are then detected using MediaPipe. Camera parameters are defined, assuming a focal length to the image width and the optical center at the image center. OpenCV functions are used to map 3D points to 2D image coordinates, generating rotation and translation vectors that describe face orientation. From these vectors, a rotation matrix is created to calculate angles along X, Y and Z-axes. Testing demonstrated robust performance both in controlled and complex conditions. Accuracy decreased for partially occluded or heavily rotated faces, but these cases resulted in undetectable faces, in line with system design.

4.4.2.3 Pretrained neural network

Another approach involved a pretrained neural network, in this case HopeNet (Ruiz et al., 2018), chosen for its accuracy and widespread use. This solution is widely used for head pose estimation in various tasks such as human-computer interaction and motion tracking. The implementation process is relatively simple. The repository is cloned, the dependencies installed and the pretrained model placed in the appropriate location. Following minor code adjustments and issues resolution, the model was run using the provided function. Modifications allowed to obtain pitch, yaw and roll values, representing the face’s deviation from frontal pose. Evaluation revealed that the model achieves high precision in estimating head rotation angles under diverse conditions. The main limitation is a noticeably longer processing time compared to other methods, due to computational demands.

4.4.2.4 Method selection

Based on the conducted analysis, the strengths and weaknesses of each approach have been identified. 2D landmark analysis provides fast processing but is prone to many critical errors. In contrast, pretrained neural network delivers very high precision but at cost of the processing time, being much longer due to computational requirements. Consequently, 3D facial landmark mapping was selected as the optimal solution, combining the advantages of both previous techniques: efficient processing time and high accuracy.

Implementation follows the previously described procedure. After obtaining the angles, tests determined that an absolute value of 0.3 is the threshold beyond which the deviation from a frontal position is excessive. Therefore, the quality score for each component (yaw, pitch, roll) is computed as shown in Equation 20:

Q C = 100 − 30 | x | 0.3 f o r | x | < 0.3 , 70 e − 4 | x | − 0.3 f o r | x | ≥ 0.3 (20)

4.5 Moments of luminance distribution

Luminance relates to the highly variable aspect of any identification process - light. Shadows, overexposure or contrast can excessively affect image quality and, therefore, the ability to recognize the subject. In this case, four separate components will be analyzed: mean, variance, skewness and kurtosis.

4.5.1 Standard compliant implementation

A luminance histogram will be used, allowing analysis of light distribution within the image. Calculations to obtain the histogram begin by converting the RGB channels to luminance and then scaling the obtained values to the range of 0–255. Pixels corresponding to each luminance level are then counted and the histogram is computed by normalizing the number of pixels to the total number of pixels in the image.

4.5.1.1 Mean

Both overly bright and underexposed facial images may negatively affect face recognition. This component will assess the deviations of the mean from the median of possible values, allowing evaluation of correct exposure by identifying under- and overexposed regions. The previously calculated luminance histogram will be used in the computations. The mean luminance is obtained from Equation 21:

h ̄ = ∑ i = 0 M h i i M (21)

Using the calculated mean, the quality measure value can be determined, penalizing both very dark and bright dark images. This was achieved using Equation 22:

Q C = ⌈ 100 SIGMOID h , 0.2 , 0.05 1 − SIGMOID h , 0.8 , 0.05 ⌉ (22)

4.5.1.2 Variance

The second luminance component refers to image contrast. It evaluates whether the image is evenly illuminated or has highly variable lighting. A high variance may indicate over- and underexposed regions, causing uneven lighting, whereas a low variance suggests subdued image details and low sharpness. Luminance variance is calculated using h ̄ computed previously as shown in Equation 23:

V A R h 0 , … , h M = ∑ i = 0 M h i i M − h ̄ 2 (23)

To calculate the quality measure value, which reflects whether the contrast is suitable, (Equation 24) is applied:

Q C = ROUND 100 ⁡ sin 60 VAR 60 VAR + 1 π (24)

4.5.1.3 Skewness

The third component will reflect the asymmetry in the grayscale distribution of the image. It enables evaluation of illumination uniformity by identifying asymmetrical regions that are either too dark or too bright, which could influence recognition accuracy.

Skewness is computed using σ = VAR, where VAR was obtained previously. Then we use Equation 25:

γ h 0 , … , h M = ∑ i = 0 M h i i M − h ̄ 3 σ 3 (25)

Using the obtained value we can compute the skewness quality score using Equation 26:

Q C = ROUND 100 1 − SIGMOID γ , 2,0.35 (26)

4.5.1.4 Kurtosis

The last component of the luminance distribution assesses kurtosis, indicating regions that are excessively dark or bright relative to the whole image. To compute luminance kurtosis, σ = VAR is used, with VAR determined previously. It is then substituted into Equation 27:

κ h 0 , … , h M = ∑ i = 0 M h i i M − h ̄ 4 σ 4 (27)

The quality score for the luminance kurtosis, is calculated using Equation 28:

Q C = ROUND 100 SIGMOID κ , 2,0.5 1 − SIGMOID κ , 10,2 (28)

4.5.2 Custom implementation

For comparison, two luminance analysis approaches have been picked, differing in methodology and implementation. The first uses popular libraries that provide a wide range of functions for image processing and statistical calculations. The second approach uses classical histogram analysis to map the pixel intensity distribution in the image.

4.5.2.1 Using OpenCV, NumPy and SciPy libraries

Using libraries designed for accurate image processing and mathematical computation such a OpenCV, NumPy and SciPy allows us to examine four key parameters of luminance: mean, variance, skewness and kurtosis. Implementation is straightforward due to the intuitive nature of the tools. Initially, the image is converted to grayscale and the 2D pixel matrix is flattened into 1D array to simplify calculations. Using the prepared image, we deploy built-in functions from NumPy to calculate the required luminance moments. Although, such method may seem simple, it remains highly effective. Its reliability comes from the precision of the libraries involved, which are recognized as industry standards.

4.5.2.2 Computing the luminance histogram

The implementation begins by converting the image into grayscale, simplifying brightness analysis. Then, a histogram of pixel intensities is calculated across 256 brightness levels, allowing a detailed image structure assessment. From this data, various statistical parameters are derived: the mean representing the average brightness; variance measuring the spread of intensity values; skewness indicating the asymmetry of the brightness distribution; and kurtosis evaluating the presence of extreme brightness regions compared to the average lighting. Similar to the previous method, we can achieve high accuracy with simplicity.

4.5.2.3 Method selection

Conducted tests indicated a similar effectiveness of both of analyzed methods, both in terms of the results and execution time. Therefore, the final approach was dictated by implementation convenience. Since the widely used libraries (OpenCV, NumPy, SciPy) are more intuitive and faster to implement, they allow for overall higher process efficiency. Additionally, considering the popularity of these libraries, numerous resources can be found on community forums regarding problem-solving.

Each component was implemented according to the previous description. In order to obtain quality measure values for each luminance component, it was necessary to determine which parameter ranges define good image quality. For this purpose, an extensive literature review and a series of tests were conducted, yielding ranges corresponding to proper image illumination. After defining these ranges, the quality measure values for each component were calculated as follows.

The ideal mean luminance value was set at 150, with the range (120, 180) representing satisfactory quality. Within this range, QC should decrease gradually, while beyond the decline should be steep, as shown in Equation 29:

Q C = 100 − 1 3 150 − mean for 120 ≤ mean ≤ 150 , 100 − 1 3 mean − 150 for 150 < mean ≤ 180 , 90 − 5 mean − 120 for mean < 120 , 90 − 5 mean − 180 for mean > 180 . (29)

The ideal variance value was set to be 10,000, with the acceptable range (5,000, 15,000) where the QC decreases gradually. Outside this range, the value decreases rapidly, as shown in Equation 30:

Q C = 100 − 1 500 10000 − variance for 5000 ≤ variance ≤ 10000 , 100 − 1 500 variance − 10000 for 10000 < variance ≤ 15000 , 90 − 1 200 5000 − variance for variance < 5000 , 90 − 1 200 variance − 15000 for variance > 15000 . (30)

For both skewness and kurtosis, the ideal value is 0. Acceptable quality corresponds to the interval (−0.5, 0.5), where the QC decreases gradually. Outside this range, the value decreases rapidly, as shown in Equation 31:where T is the value of skewness or kurtosis respectively.

Q C = 90 + 20 T + 0.5 for − 0.5 ≤ T ≤ 0 , 100 − 20 T for 0 < T ≤ 0.5 , 90 − 50 T + 0.5 for T < − 0.5 , 90 − 50 T − 0.5 for T > 0.5 . (31)

5 Experiments, optimization and results

The experiments were structured into two stages in order to ensure a systematic and comprehensive evaluation of the proposed solution. The first stage focused on optimizing the image quality assessment framework itself. It involved tuning the weights for each measure within the overall scoring and determining the optimal decision thresholds for both developed systems. The goal was to analyze the contribution of each quality measure to the final assessment score and to identify the configuration that provides the best evaluation of facial image quality.

The second phase of the experiments analyzed the performance of the optimized quality assessment framework in a face recognition scenario. In this phase, the relationship between estimated image quality and recognition performance was examined to verify whether both systems reflect biometric usability. This two-phase experimental design enabled both a detailed analysis of individual system components and a holistic assessment of their impact on face recognition accuracy.

5.1 Image quality measures testing and tuning

Tests were conducted in two main categories, allowing for detailed evaluation of the implemented measures. The first category focused on analyzing the impact of weights assigned to individual measures on the final quality assessment score, identifying the most influential components and their effect on evaluation accuracy. The second category examined the modification of the three-level quality scale thresholds (bad, good, very good). For the testing procedure a dataset of facial images was created, categorized according to quality. Images were sourced from public datasets: Lagenda (Kuprashevich and Tolstykh, 2023) and LFW (Labeled Faces in the Wild) (Huang et al., 2007), selected to be highly diverse, considering both the quality evaluated by each measure and the people depicted, including different age and ethnic groups. Each image was labeled according to its quality category. To evaluate the impact of weight and threshold variations on system performance, metrics including FPR (False Positive Rate), FNR (False Negative Rate) and EER (Equal Error Rate) were applied. In our case the FPR denotes cases in which an image of bad quality is incorrectly classified as good or very good quality, as well as cases in which an image of good quality is misclassified as very good quality. Conversely, the FNR refers to situations where an image of very good quality is incorrectly classified as bad or good quality, or where an image of good quality is misclassified as bad quality. These metrics allowed for a precise analysis of how efficiently the system works.

As an illustration of our reasoning process, we present the procedure used to derive the weights for individual quality measures, using the eyes open metric as an example. The same methodology is consistently applied to all remaining metrics.

For the standard based methodology, the quality measure was implemented in accordance with the recommendations 4.3.1. In the case of the custom implementation, three distinct methods described in the corresponding Section 4.3.2 were first implemented and then subjected to extensive performance evaluation. Each method was tested on carefully selected datasets containing images of subjects with open, closed, and partially closed eyes, as well as datasets diversified with respect to race and age. The performance of each method was thoroughly assessed based on error-rate analysis and a visual inspection of their operational behavior. Based on the computed performance metrics and the conducted analysis, the most effective method was selected and incorporated into the final implementation.

After conducting the analysis and implementing the ultimately selected methods in the final system, it was necessary to determine the weights of the individual quality measures, which are required to compute the overall score. Tests were carried out on the Lagenda dataset (see Kuprashevich and Tolstykh, 2023) by varying the weight assigned to the investigated measure. Specifically, the weight of the target measure was gradually increased, while the weights of all remaining measures were adjusted proportionally so that the resulting score would reflect only the contribution of the target measure. For example, when testing a weight of 0.6 for a given measure, the remaining four measures were assigned equal portions of the remaining weight, i.e., 0.4 divided by 4, resulting in 0.1 per measure. The effect of varying the weight of the eyes open measure on the error rates is illustrated in Figure 3.

5.1.1 Measure weight analysis

The evaluation was performed both for ISO standard and custom implementation, following identical set of steps. Each measures’ weight was incrementally adjusted from 0 to 1 in 0.1 steps. Initial thresholds were applied to convert the numeric scores into the descriptive scale, enabling the comparison of final result with dataset labels and computation of system effectiveness.

When looking at the results in Table 1, we can clearly see the differences in how much importance is given to each measure. In the ISO compliant implementation, the presence of single face had the strongest influence, meaning that the system detected faces very accurately. The head position and eye openness had equal importance showing their equal contribution to the final score. The natural skin color was given a lower weight, possibly due to reduced detection accuracy under poor lightning condition. Meanwhile, the luminance distribution was considered the least important factor, also ranking low in the custom measures. This is related to the limited correlation between luminance values and overall image quality, as luminance is a complex and highly variable component influenced by many external factors. In the custom measures, the natural facial color was assigned the highest weight, indicating that the implementation effectively meets the intended requirements. The lowest weight was given to the presence of a single face, suggesting possible detection inaccuracies. The remaining weights are distributed gradually, with open eyes receiving a high value—confirming the measures’ effectiveness, while face position holds moderate importance, as slight head tilts do not significantly affect recognition in biometric systems.

TABLE 1. Table showcasing the best weight values, optimized for ISO compliant and custom implementation.

5.1.2 Final threshold analysis

After determining the optimal weights for each measure, the previously assumed thresholds responsible for converting the numerical output into descriptive scale were verified. The process was divided into two parts–determining the lower and upper threshold. Using the fixed optimal weights, a final score was computed for all samples, while the decision threshold was systematically incremented by 2%. For each threshold value, the results were compared with ground-truth labels. Based on these comparisons, the FPR and FNR metrics were calculated for each threshold - Figures 4, 5. The resulting data allowed for detailed analysis of specific threshold impact on system performance.

Analyzing the selected threshold values - Table 2, we can see the differences in the thresholds assigned to each implementation. For the ISO compliant implementation, the gap between the lower and upper thresholds is relatively small, indicating a smaller range that defines good image quality. In contrast, in the custom implementation, the smallest range corresponds to bad quality. This results from the differences in the weights of individual measures, which significantly influence how images are assigned to specific categories. The thresholds vary depending on how restrictive the selected weights are, and vice versa. Interestingly, the upper threshold value is similar for both implementations, which may indicate comparable sensitivity when distinguishing between good and very good images. From the perspective of optimization, it is therefore crucial to select complementary weights and thresholds to ensure a balance between sensitivity and the stability of the system’s operation.

TABLE 2. Table showcasing best values for the thresholds, optimized for ISO compliant and custom implementation.

5.1.3 Results

After adjusting the weights and thresholds, the optimized programs were evaluated for performance and accuracy in image quality assessment. The final metrics are listed in Table 3 and the classification confusion matrix is presented in Figure 6. In our case the FPR denotes cases in which an image of bad quality is incorrectly classified as good or very good quality, as well as cases in which an image of good quality is misclassified as very good quality. Conversely, the FNR refers to situations where an image of very good quality is incorrectly classified as bad or good quality, or where an image of good quality is misclassified as bad quality.

TABLE 3. Table showcasing the values of normalized classification error of image quality, optimized for ISO compliant and custom implementation.

The results indicate that the custom implementation achieved approximately 8% higher efficiency across both indicators, reflecting the advantages of our solutions. In contrast, the ISO-compliant methods relied on outdated approaches which struggled to perform well under challenging evaluation scenarios. The modernized solution proved to be more effective and accurate, adapting well to diverse visual conditions.

5.2 Image quality assessment’s impact on face recognition performance

This part evaluates the effect of image quality assessment on the performance of a face recognition system. The dataset used (see VGGFace2 Dataset, 2018) includes 10 images per identity, including 200 identities, ensuring variability. Initially, the reference recognition performance is measured using all available images without discarding any samples in order to obtain reference face similarity threshold.

For the development of our face recognition system, we evaluated several state-of-the-art models and selected AdaFace (Kim et al., 2022), as it demonstrated high accuracy and reliability across diverse datasets while remaining straightforward to implement. This enables robust performance across varying conditions such as low resolution or occlusion. While other recent models reported high benchmark accuracy, they posed practical challenges in implementation, requiring complex preprocessing or specialized training procedures. AdaFace provided a balance of competitive accuracy, robustness, and ease of integration, making it the most suitable choice for our application.

A face recognition system AdaFace (Kim et al., 2022) was applied to the dataset, and a one-to-many evaluation protocol was adopted. In this setup, each image of a given identity (probe) was compared against all other images of the same identity (gallery), capturing the full intra-class variability. Simultaneously, each probe image was compared against all images of other identities to account for inter-class discriminability. Cosine similarity between embeddings was used to generate similarity scores for all probe-gallery and probe-impostor pairs. From these aggregated scores, the Receiver Operating Characteristic (ROC) curve was computed–Figure 7. The baseline decision threshold of 0.109, corresponding to an Equal Error Rate (EER) of 0.114, was selected from the ROC curve and serves as the reference for all subsequent analyses.

The ROC curves (Figure 7) show a steep rise at low false match rates, indicating strong separation between genuine and impostor score distributions. This means that most genuine matches are correctly accepted before false matches increase, demonstrating the high discriminative capability of the AdaFace (Kim et al., 2022) embeddings under a one-to-many evaluation protocol. The Equal Error Rate occurs at a relatively low false match rate, reflecting a balanced trade-off between false acceptances and false rejections. The log-scale ROC further highlights performance in the low-FMR region, which is critical for security-sensitive face recognition applications.

5.2.1 Experiments

In the experimental phase, images are progressively removed from the VGGFace2 dataset (see VGGFace2 Dataset, 2018) based on their quality scores. For each iteration, all images with a quality score below a given threshold are excluded, and the face recognition process is repeated on the remaining fraction of the images. For mated comparisons, every image is compared against all other images of the same person (one-to-many comparisons) to fully capture intra-class variability. The quality threshold is increased in increments of 5%, and for each threshold step, FNMR and FMR are recalculated. Simultaneously, the proportion of rejected images is recorded. This iterative removal simulates the effect of discarding low-quality images, and if the quality measure is predictive of recognition performance, FNMR is expected to decrease as the threshold increases, while FMR may remain stable or also decrease depending on the dataset characteristics.

The plots in Figure 8 illustrate the impact of progressively removing low-quality images on recognition performance. As the proportion of rejected images increases, FNMR shows a clear decreasing trend, indicating improved reliability of genuine matches when poor-quality samples are excluded. In contrast, FMR remains largely stable across most rejection levels, with only minor variations at high rejection rates, suggesting that quality-based filtering primarily reduces false non-matches without adversely affecting impostor discrimination.

5.2.2 Results

The effect of image quality on the performance of the AdaFace (Kim et al., 2022) face recognition system was evaluated using the one-to-many comparison protocol. For each predefined quality threshold, all images with a quality score below the threshold were excluded from the dataset, and the face recognition process was repeated on the remaining images. FNMR and FMR were recalculated at each step, and the proportion of discarded images was recorded to quantify the impact of progressive pruning.

The one-to-many similarity computation used in this experimental setup served as a robust and informative measure of recognition performance, as it effectively captured variations in image quality and intra-class variability by comparing each probe image against all other images of the same identity.

Notably, the steeper FNMR reduction observed for the custom implementation compared to the ISO-based approach suggests a stronger correlation between the assigned quality scores and actual recognition performance.

When applying the ISO landmark-based quality measure, many images in the previously considered FaceScrub (FaceScrub Dataset, 2014) dataset were assigned a quality score of zero, primarily due to the cropped nature of the faces, which prevented the algorithm from detecting a sufficient number of landmarks. In contrast, the custom quality assessment implementation used in this study did not encounter similar issues, as the face detection process proved more robust to partial cropping and variations in alignment. This suggests that the face detector integrated in the custom pipeline is less sensitive to incomplete faces, allowing reliable similarity computation even when images are partially cropped. Consequently, ISO-based filtering had limited effect on the FNMR and FMR curves, whereas the custom approach maintained consistent performance across the dataset.

Overall, these results demonstrate that removing low-quality images can substantially improve recognition performance, particularly when using a quality measure that effectively reflects embedding reliability. The combined ROC and Error vs. Discard Characteristic analyses further highlight that robust quality assessment enables the system to operate in favorable low-FMR regions while achieving lower FNMR. The superior behavior observed for the custom quality assessment underscores the importance of selecting modern, state-of-the-art quality measures that go beyond strict standard compliance, ensuring consistent and reliable performance in practical face recognition scenarios.

6 Conclusion

As a part of this work, two competing systems for assessment face image quality in biometric systems were developed. Both of them utilized measures defined in the emerging ISO/IEC 29794-5:2023 norm. One implementation strictly followed the norm’s recommendations while the other utilized recent state-of-the-art techniques in image analysis. Both systems allowed for flexibility to define custom measure weights and threshold values. The purpose of this work was not only to design these quality assessment systems, but also to thoroughly test and compare their performance. Accordingly, both implementations were optimized in terms of all possible parameters, using extensive testing performed on the created dataset. The experiments confirmed the hypothesis, proving higher efficiency of state-of-the-art solution which outperformed the ISO-based system. While the standard defines a robust framework, it relies on outdated algorithms which do not fully align with current technological demand.

Thanks to the NIST FRVT ³ , there are known top-performing state-of-the-art solutions for image quality assessment. However, the vast majority of the submitted algorithms are proprietary and thus not available for public use. Keeping this in mind, it is reasonable to search for open/free solutions that perform better than the reference one, and are available at a low cost.

Implementing image quality assessment systems in biometric applications is crucial for the effective operation of both commercial and industrial sectors. Quality analysis contributes to precise error mapping, therefore improving explainability and overall transparency. This is particularly important in high-risk environments such as finances, where unauthorized access could lead to legal actions requiring a full reconstruction of system decision process. In addition, automated quality enhances system’s accuracy with relatively low operational cost. Such technology proves to be especially valuable in sectors relying on fast, accurate and self-sufficient user verification, such as security, banking or public services.

6.1 Limitations of the study

Despite the systematic approach and careful design of the experiments, the study has several limitations. First, the datasets used were relatively small and may not fully reflect real-world variability. The primary dataset consists of 300 manually annotated images divided into three quality subsets, which, while sufficient for initial analysis, may not capture all variations in lighting, pose, demographic factors, or capture devices. The evaluation dataset, based on VGGFace2, includes 2000 images from 200 subjects, providing more diversity but still not fully representing operational scenarios.

Second, the study focuses on only five selected quality measures from ISO/IEC 29794-5:2023, chosen based on their typical presence in eKYC selfies for high-risk financial applications. While relevant for this context, these measures do not cover all known metrics, and the selection process may introduce bias. Not all methods for computing each measure were implemented, which may limit the completeness of the analysis and the comparability with other quality assessment systems.

Finally, the evaluation relies on a single face recognition system, AdaFace, and the performance was analyzed using approach similar to the Error vs. Discard Characteristic. While this enables controlled testing, the results may not fully generalize to other algorithms, especially those with different sensitivities to image quality. The iterative thresholding procedure also assumes that measures independently predict recognition performance, which may not always hold. Overall, these limitations suggest that the results indicate general trends rather than definitive performance guarantees.

Correction note

This article has been corrected with minor changes. These changes do not impact the scientific content of the article.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Andrejs Romanovs, Riga Technical University, Latvia

Reviewed by: Ankita Guleria, Panjab University, India

Joko Triloka, Informatics and Business Institute Darmajaya, Indonesia

Footnotes

During our investigation of the topic, the work on the standard was still in progress, so our solutions were based on the newest available reference document ISO/IEC 29794-5:2023 (CD3) (International Organization for Sta ndardization and International Electrotechnical Commission, 2023).

https://pages.nist.gov/frvt/html/frvt_quality.html

https://pages.nist.gov/frvt/html/frvt_quality.html