Background: Facial recognition systems utilizing deep learning techniques can improve the accuracy of facial recognition technology. However, it remains unclear whether these systems should be available for patient identification in a hospital setting. Methods: We evaluated a facial recognition system using deep learning and the built-in camera of an iPad to identify patients. We tested the system under different conditions to assess its authentication scores (AS) and determine its efficacy. Our evaluation included 100 patients in four postures: sitting, supine, and lateral positions, with and without masks, and under nighttime sleeping conditions. Results: Our results show that the unmasked certification rate of 99.7% was significantly higher than the masked rate of 90.8% (p < 0.0001). In addition, we found that the authentication rate exceeded 99% even during nighttime sleeping. Furthermore, the facial recognition system was safe and acceptable for patient identification within a hospital environment. Even for patients wearing masks, we achieved a 100% success rate for authentication regardless of illumination if they were sitting with their eyes open. Conclusions: This is the first systematical study to evaluate facial recognition among hospitalized patients under different situations. The facial recognition system using deep learning for patient identification shows promising results, proving its safety and acceptability, especially in hospital settings where accurate patient identification is crucial. 2024 https://creativecommons.org/licenses/by/4.0/ Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license ( https://creativecommons.org/licenses/by/4.0/ ). Background: Facial recognition systems utilizing deep learning techniques can improve the accuracy of facial recognition technology. However, it remains unclear whether these systems should be available for patient identification in a hospital setting. Methods: We evaluated a facial recognition system using deep learning and the built-in camera of an iPad to identify patients. We tested the system under different conditions to assess its authentication scores (AS) and determine its efficacy. Our evaluation included 100 patients in four postures: sitting, supine, and lateral positions, with and without masks, and under nighttime sleeping conditions. Results: Our results show that the unmasked certification rate of 99.7% was significantly higher than the masked rate of 90.8% ( p < 0.0001). In addition, we found that the authentication rate exceeded 99% even during nighttime sleeping. Furthermore, the facial recognition system was safe and acceptable for patient identification within a hospital environment. They have been employed in settings other than hospitals [ 2 , 3 ], e.g., in security cameras [ 4 ], examinations for immigration [ 5 ], and entrance and It is well known that wearing masks reduces face recognition accuracy, as with iPhones and other devices [ 10 , 11 ]. However, according to a recent study of masked face recognition [ 12 ], authentication is possible regardless of the type or position of the mask [ 13 ]. Facial recognition is now expected to play an important role in future healthcare systems. The adoption of facial recognition systems in hospitals is expected to reduce human error and prevent patient misidentification. Wrong-patient errors occur in all steps of diagnosis and treatment in hospitals, highlighting the need to improve the accuracy of patient identification. For facial recognition in hospitals, it is necessary to consider conditions such as lying on a bed or dimmed lighting. The illuminance required for face recognition is 200 lux or higher. Silverstein showed that when imaging the same face in different light levels from 60-285 lux, the process is less accurate in lower light and only captures consistent face data if the ambient light is sufficient [ 8 ]. Munn and Stephan pointed out that facial recognition performance declines when a person is lying down because of the physical change in their facial expression by gravity [ 26 ]. Under the nighttime sleeping condition, we successfully authenticated 99.3% (298/300) of the patients in either the left or right lateral position without a mask, in low light, and with closed eyes ( Figure 5 ). 4. Discussion In this study, we verified a facial recognition algorithm for hospital settings and obtained three essential findings. (1) The unmasked certification rate of 99.7% was significantly higher than the masked rate of 90.8%. For patients not wearing masks, the deep learning-based facial recognition system showed 99.7% accuracy even under adverse conditions (low illumination, eyes closed, supine position), excluding the normal condition. (2) Furthermore, the system showed 99.3% accuracy under nighttime sleeping conditions (supine position or lateral position, low illumination, closed eyes) without a mask. (3) For patients wearing a mask, the system showed 100% authentication accuracy if the patient was sitting with their eyes open. Previous studies on facial recognition have not been conducted among outpatients or inpatients in medical settings. This is the first study to perform facial recognition in different situations with combined conditions, e.g., the sitting position, supine position, under low lighting, or with eyelids closed, representing a hospital ward at night. The satisfactory results regarding the nighttime sleeping conditions in this study mean that we can use the system in a hospital room during actual nighttime hours. Recently, algorithms has improved and infrared 3D facial recognition has evolved to not reduce the authentication rate even in low illumination [ 27 ]. Facial recognition already had some advantages (i.e., the system does not need direct contact, works distantly, and does not rely on the patient’s response).

Facial recognition technology has become increasingly everywhere in various domains, from security and surveillance to personal device authentication. However, its effectiveness can be significantly hindered in low-light conditions, where images often lack sufficient illumination for accurate recognition. This study proposes a novel approach to enhance facial recognition accuracy in low-light conditions using Convolutional Neural Networks (CNNs), Deep Retinex Decomposition Network (DRDN), and CenterFace algorithm. The methodology leverages CNNs for robust feature extraction, while DRDN corrects illumination by decomposing images. CenterFace integrates feature fusion and denoising layers for discriminative facial features and noise mitigation. Experimental results demonstrate a remarkable improvement in recognition performance, exceeding 80% accuracy. This approach showcases the potential of CNN-based methods with advanced techniques to enhance reliability in real-world facial recognition applications, particularly in low-light environments.

Prior research has shown that under the illumination controlled and user cooperative conditions, nearly all of face recognition method perform very well, but when it comes to the variant illumination, pose and expression (PIE) conditions, the performance of these methods degrade seriously. Face recognition in variant PIE conditions is one of the most challenging problem in this field. In this paper, a method based on SIFT features is used to research on face recognition in variable PIE conditions. Three professional PIE-variable face database are used to experiment the performance of SIFT method. The experiment results show the huge potential of SIFT method in application to face recognition in variable PIE conditions. Introduction During the past several decades, face recognition has been shown great potential applications in public security, law enforcement and surveillance, digital entertainment, access control and others. Considerable progress has been made in thid field [1]. Now in controlled conditions, face recognition systems perform very well, but when it comes to the variant illumination, pose and expression (PIE) conditions, the performance of systems need to be improved.Some pointed methods are attempted to investigate the problem of face recognition in variable illumination,pose and expression conditions respectively.[2] has shown the attempt to reduce the influence of light source or a large amount of training data or some 3D face models whose facial shapes and albedos are obtained in advance, which is not practical for most real word scenarios.The 3D model methods based on illumination modeling are investigate to reduce the influence of illumination factor[3].A same problem of these method is that the computation is not efficient enough.When it comes to the adverse factors of pose and expression, an author had proposed an approach that couples the geometric information of face with the texture one[4]. In paper [5], a novel perception inspired non-metric partial similarity measure was introduced, which could help capturing the prominent partial simlilarities that were dominant in human perception. The authors proposed two methods to automatically set the similarity threshold, which were based on the general golden section rule and the maximum margin criterion. In the actual application of face recognition systems, these adverse factors are almost concurrent, so an effective method which is applicable to the PIE variant conditions is needed to be proposed. In this paper, the Scale Invariant Feature Transform (SIFT) is proposed to dealing with above problem. SIFT features are extracted from images to help in reliable match between different views of the same object by SIFT. These features are invariant to scale and orientation, and highly distinctive of the image [6].Paper [7, 8] and [9] have shown the application of SIFT for recognizing faces in controlled situations, and performed well. This paper proved that SIFT could also perform well under variable PIE situation by some experiments based on several professional PIE-variable face database. The rest of this paper is organized as follows. SIFT method is reviewed. Next, the application of SIFT are described, to extract SIFT features of PIE-variant faces images. Then, extensive experiments are conducted and results are analyzed. Finally, Analysis and conclusions are presented. International Conference on Information Sciences, Machinery, Materials and Energy (ICISMME 2015) © 2015. The authors Published by Atlantis Press 688 The Scale-invariant Feature Transform The scale invariant feature transform, called SIFT descriptor, has been proposed by Lowe [6], and proved to be invariant to image rotation, projective transform, scaling translation, and partly illumination changes. So the features extracted by SIFT have been shown to be invariant to image rotation and scale and robust across a substantial range of affine distortion, addition of noise, and change in illumina