The COVID-19 pandemic ushered in widespread mask mandates across many nations to curb virus transmission. Masks, shown to be cost-efective safeguards in healthcare settings, are poised to remain a societal norm. However, their use substantially obscures facial expressions, complicating emotion identification. Hence, assessing the impact of masks on facial expression recognition and emotional interpretation is imperative. Previous studies identified accuracy losses and increased confusion in emotion recognition with masks but didn't explore underlying causes. This study delves into these confusions by analyzing facial characteristics influencing each expression, using the LIME explainable AI technique. Starting with Faigin's facial feature definitions for expressions, we group similar expressions sharing visible features under masks, hypothesizing they're recognizably similar. We employ a CNN model on a masked facial expression dataset to test our hypothesis.
The advent of the COVID-19 pandemic led to the widespread use of face masks to reduce virus transmission. While originally a public health measure, masks have proven to be a cost-efective mechanism for personal protection, particularly in healthcare settings, and are likely to remain common in many environments. Moreover, facial occlusion is not exclusive to pandemics: it also occurs in everyday situations such as the use of protective gear, outdoor activities, or cultural and religious practices. In all these cases, the lower part of the face—including the chin, mouth, and nose—is partially or fully hidden, making emotion recognition more challenging. Understanding how this occlusion afects emotional perception is therefore critical for the design of inclusive and afective interactive systems.
There is evidence in the literature that wearing a mask negatively impacts the ability to discern
between diferent facial expressions as it causes the loss of a significant amount of facial information
[
Several studies have examined this issue in more depth. Carbon [
CEUR
ceur-ws.org
In this study, we aim to go a step further by analyzing which visible facial features contribute to systematic confusion when a person wears a mask. We hypothesize that expressions which share the same visible features when masked are likely to be confused. Previous studies have shown that Convolutional Neural Networks (CNNs) can detect subtle facial features more efectively than human observers, often leading to more accurate classification of expressions [ 13, 14]. To test our hypothesis, we train a CNN model on a dataset of masked facial expressions and interpret the resulting classifications using an explainable artificial intelligence technique (XAI) called Local Interpretable Model-agnostic Explanation (LIME).
In order to analyze how wearing a face mask may afect the recognition of facial expressions, we begin by identifying the facial features that define each expression, based on the theoretical framework proposed by Faigin [15], who divides the face into three main zones: the upper part of the face (including forehead and eyebrows), the eyes, and the mouth. He defines similarity between facial expressions based on the configuration of facial elements within these zones. Two expressions are considered “similar” when they share configurations in one or more of these regions. Taking this definition as our baseline, we group facial expressions into sets of “similar facial expressions”, where each set includes expressions that share the same visible facial features when a mask is worn (i.e., primarily those in the eye and upper face regions). Accordingly, we hypothesize that all expressions within the same set are likely to be recognized as the same facial expression, given the limited visibility caused by the mask.
From the perspective of Human-Computer Interaction (HCI), this research provides valuable insights into how facial occlusion interferes with afective computing. The findings are applicable to various interactive systems where emotional awareness is essential, such as virtual assistants, e-health, or socially assistive robots.
The paper is structured as follows: Firstly, we introduce the UIBVFED-Mask dataset, which will be employed extensively in our study. Subsequently, we formulate hypotheses regarding facial expression confusion when wearing face masks for each emotion. In Section 4, we provide an overview of our neural network model, data preprocessing procedures, and our prediction explanation methodology. Section 5 is dedicated to present our findings and engaging in discussions for each emotion category. Lastly, we summarize our conclusions.
For the experiments, the UIBVFED-Mask dataset [16] is used. The UIBVFED-Mask dataset is an extension of the UIBVFED dataset [17]. UIBVFED is a database formed by virtual characters performing 32 facial expressions, classified based on the six universal emotions according to Gary Faigin ( Anger, Disgust, Fear, Joy, Sadness, and Surprise) [15] plus the Neutral emotion. The UIBVFED-Mask dataset comprises the same images as the UIBVFED dataset but has been reconstructed to include face masks. Figure 1 shows an example of the Neutral expression in the original UIBVFED dataset and the corresponding image in the UIBVFED-Mask dataset, where we can clearly see the portion of the face hidden by the mask.
The images of the facial expressions that compose the dataset were generated following the guidelines of the Facial Action Coding System (FACS) [18]. So, the deformations applied to the 3D models have a direct correspondence with the Action Units (AUs) that are associated to each expression. This procedure ensures that images are labeled objectively. Moreover, the usage of synthetic datasets has actually proven to be a good replacement for real image datasets since they achieve recognition rates that are comparable to the real ones [19, 20].
The dataset is comprised of a total of 660 images, with 20 avatars representing 33 diferent facial expressions each. Table 1 displays the distribution of images per emotion.
(a) (b)
Despite all the advantages of synthetic datasets, the UIBVFED-Mask dataset has associated some limitations that may lead to some emotion’s confusion: • Geometry: Facial geometry of the characters can make certain values, for example those that difer in intensity, indistinguishable from each other. In addition, the pressure of the eyes is produced by a deformer that acts on the lower eyelid. The most visible efect of this are the characteristic wrinkles in the lower and exterior parts of the eyes. The UIBVFED-Mask models do not have enough geometric density to reflect these wrinkles in detail. For this reason, this feature can cause confusion. • Texture: The UIBVFED-Mask characters have the eyebrows defined by texture, not geometrically.
As in reality, depending on the shape of the eyebrow or the makeup of the character, a relaxed eyebrow may look like a furrowed brow. Therefore, certain characters can lead to confusion. • Facial animation: Facial animation of the characters is based on blendshape deformations that afect the geometry of the skin and the lower denture. These deformations originate from the character configuration in the Autodesk Character Generator software application, and they are not inheritable by hierarchy. Therefore, the mask geometry is not afected by the animation. This might seem an inconvenience since the image will not have this information. In any case, in this study we do not consider image sequences where the movement of the mask deformation could be more helpful, but we only analyze facial expressions at their zenith.
In accordance with Faigin [15], facial expression recognition depends on the role of the face muscles. In his work, the author focuses on the action of muscles in three key areas of the face: (1) the forehead and brows; (2) the eyes; and (3) the mouth and chin. The same author empathizes that “an expression will only be clear and unambiguous when there is action in both the eyes/brow and the mouth at once”.
Wearing a face mask occludes the area of the mouth and chin, that is, one of the three key areas. Therefore, it is almost impossible, according to the theory, to recognize a facial expression in a clear and unambiguous way. So, confusion between facial expressions is to be expected when wearing a face mask, that is to say, when the area of the mouth and chin is occluded.
In order to evaluate how wearing a face mask may afect the recognition of a facial expression, in this section the facial expressions described by Faigin are analysed. This analysis focuses on the features of the upper part of the face -i.e., the key areas of the face that remain visible when wearing a face mask: (1) the forehead and brows, and (2) the eyes.
While Faigin describes in detail the muscle movements and visible features associated with each facial expression, he does not address the impact of partial facial occlusion –such as the use of face masks– on expression recognition. The hypothesis proposed in this work applies Faigin’s definitions to the masked context, aiming to predict which expressions may become visually indistinguishable when the lower face is hidden.
In the following sections, for each of the six universal emotions plus the Neutral emotion, we present a table summarizing, for each of its associated facial expressions according to Faigin [15] (first column), the facial features of the two key areas of the upper part of the face (second column). The facial expressions associated to the same emotion that share all their visible facial features are shown in the same row, meaning that they are plausible to be confused. Besides, for each row, we present (in the third column) the number of similar facial expressions (we define “similar” in terms of sharing the same facial features) associated to other emotions. Similarities between diferent emotions are grouped by colour. So, facial expressions of the same colour could be associated (correctly or incorrectly) to the same emotion.
Figure 2 shows an avatar (Isabel avatar) performing the Neutral facial expression without (see Figure 2a) and with a face mask (see Figure 2b). In this figure, it can be observed that the facial features corresponding to the Neutral facial expression that are visible when wearing the face mask are relaxed eyebrows and the eyes opened without pressure. In this scenario, the information of the facial feature of the relaxed mouth is occluded with the face mask and, therefore, this information is lost. (a) (b) (c) (d)
Figure 3 shows the Isabel avatar wearing the face mask and performing the Neutral facial expression (see Figure 3a) and their similar facial expressions according to Table 2: Disgust facial expression (see Figure 3b), False Laughter 1 facial expression (see Figure 3c), and False Smile facial expression (see Figure 3d). All the images of Figure 3 present relaxed eyebrows and the eyes opened without pressure; the only subtle diference lies in how widely the eyes are opened. Observing theses images, it is almost impossible to distinguish which of them correspond to the neutral, disgust or joy emotion; and, therefore, they could be associated (correctly or incorrectly) with any of these emotions.
Table 3 summarizes the facial features of the expressions associated to the Anger emotion that are visible when wearing a face mask (see second column). In the third column, it can be observed that all the facial expressions associated to the Anger emotion share their visible facial features with one of the facial expressions of the Joy emotion: Ingratiating Smile (see Table 6).
The expected confusions for the Disgust emotion are summarized in Table 4. Besides the Disgust facial expression already analyzed in the Section 3.1, we can observe in Table 7 that the Physical Repulsion facial expression could be confused with two of the six facial expressions associated with the Sadness emotion: Crying Closed Mouth and Crying Open Mouthed facial expressions.
Figure 5 shows the Isabel avatar wearing the face mask and performing the Physical Repulsion facial expression and its similar facial expressions of Sadness emotion (according to Table 4): Crying Closed Mouth and Crying Open Mouthed facial expressions. All the images of Figure 5 present furrowed eyebrows and the eyes closed with pressure, which make them indistinguishable.
In Table 5 we can see the expected confusions associated to the Fear emotion. Two of the facial expressions associated to the Fear emotion (Terror and Very Frightened facial expressions) share their visible facial features with the only facial expression of the Surprise emotion: Surprise (see Table 8).
In previous sections, we have already analyzed the facial expressions that could be confused with the facial expressions associated with the Joy emotion (see Table 6): the False Laughter 1 and the False Smile facial expressions have already been analyzed in the Section 3.1, and the Ingratiating Smile facial expression has already been analyzed in the Section 3.2.
The expected confusions for the Surprise emotion can be seen in Table 8. This facial expression could be confused with facial expressions associated to the Fear emotion: Terror and Very Frightened facial expressions (see Table 5).
In the tables of the previous sections, similarities between diferent emotions have been grouped by color, while similarities within the same emotion have been grouped in the same row. According to these criteria, all colorless expressions should be perfectly distinguishable. Table 9 summarizes the theoretical expected confusion between emotions.
In this work we use a Convolutional Neural Network (CNN) to analyse the efects of wearing a face mask in the automatic recognition of facial expressions.
The main contribution of this work is not the CNN model, rather to test the factors that wearing a face mask introduces in both training and recognition of human emotions on facial expression images. Therefore, we don’t pretend to define a very precise neural network to recognize facial expressions, as this target has already been successfully addressed in the literature [21], but to use a simple CNN model as a baseline to analyze our hypothesis.
This section contains a detailed description of how we pre-process the data, the procedure we follow to train and test the CNN model and, finally, how we explain the predictions obtained by the model using the Local Interpretable Model-agnostic Explanation (LIME) technique.
We use a Convolutional Neural Network (CNN) following the scheme shown in Figure 7. CNNs are particularly well-suited for facial expression recognition due to their ability to automatically learn spatial hierarchies of features from image data.
As input, we have a grayscale image with a resolution of 128x128 pixels. The characteristics of the input image are extracted using a three-level combination of convolution and max-pooling. A Rectified Linear Unit (ReLU) activation function is also applied to activate only the nodes that serve our purpose. The two final layers of the CNN are fully connected layers. In the first dense layer, we flatten the last output tensor from the convolutional base and, in the final dense layer, we obtain the seven outputs corresponding to the classification classes of the emotions ( Anger, Disgust, Fear, Joy, Neutral, Sadness, Surprise).
To pre-process each image, we began by cropping the face to reduce the impact of the background on the result. Then, we converted the image to grayscale and we adjusted its resolution to fit the required dimensions of the data input to our CNN (128x128 pixels).
After completing the pre-processing step for all the images in the UIBVFED-Mask dataset, we prepared the training and testing datasets.
For the training dataset, we collected a 80% of the data, and we took the remaining 20% for the testing dataset. Both datasets contained a class distribution that was representative of the complete UIBVFED-Mask dataset (see Table 10).
We trained the CNN model previously described with the training dataset. Then, the model was tested with the testing dataset and the evaluation metrics in terms of global accuracy and confusion matrix were computed.
As a final step, and to try to obtain an explanation of the model’s outcome, we apply the LIME technique over the predictions (see Section 4.4).
We visually analyze the predictions obtained by the model using the Local Interpretable Model-agnostic Explanation (LIME) [22] technique. LIME is a widely used method to obtain local explanations of black-box models [23, 24] because it can be applied to any machine learning model due to its agnostic nature.
When explaining the classification on images, LIME depicts the main parts of the image input that contribute to the prediction using a simple approach: it perturbs the inputs of the model and observes how the new predictions behave, then it learns how the model works using a linear model through the weighting of the perturbations. The obtained explanation is not globally valid but it is locally accurate around the perturbed inputs.
What LIME highlights in their explanations are the superpixels, or collections of pixels that cover a connected area of the image, which best justify the selection of a given class. The superpixels should correspond to specific patterns of the image but normally the user can only specify the resolution of the considered areas. This poses an additional dificulty as significant features may lie within diferent superpixels. However, we can exert some control if we know the relative size of the afected areas, allowing us to adjust the number of superpixels and, consequently, their size (see Figure 8). In this study we set the number of superpixels to 50.
Nº superpixels = 25
Nº superpixels = 50
The Convolutional Neural Network model trained with face masks had an overall accuracy of 0.65. As can be seen in the confusion matrix depicted in Figure 9, we obtained perfect results for Anger (100%), very good results for Joy (89.3%) and good results for Fear (75%). Notable are the cases of Disgust and Neutral emotion, in which none of the images were correctly classified.
It is also interesting the information of the most misled emotions: Neutral, Disgust and Sadness are confused with Joy in the 100%, 91.7% and 70.8% of the cases, respectively. This confusion agrees with the theoretical emotion confusion summarized in Table 9, as some of the facial expressions associated to the emotions Neutral, Disgust, Joy and Sadness share the same visible facial features. Also the misled of 50% between the emotions Fear and Surprise is inline with the theoretical expected confusion.
Below there is a detailed discussion of the predictions obtained by emotion.
As already commented, all the images of the Neutral emotion were labeled as Joy emotion.
Theoretical confusion of emotions related to the Neutral expression, as shown in Table 2, suggests
that the Neutral facial expression (associated with the Neutral emotion) may be confused with the facial
expressions False Laughter 1 and False Smile, which are associated with the Joy emotion, as well as with
one facial expression linked to the Disgust emotion. Consequently, the behavior of the CNN model
aligns with the hypothesis we have formulated in Section 3. This observation is also consistent with
the findings presented in Carbon [
(a) Neutral (b) Neutral
In Figure 10 we can observe the results of applying LIME to the predictions of two images of Neutral emotion labeled as Joy emotion. The areas highlighted in blue correspond to the regions on which the model relies to make predictions for these expressions. In both cases, the superpixels highlighted of the images coincide with the area of the eyes and they depict the relaxed eyebrows and opened eyes without pressure.
Our trained CNN model correctly classifies all Anger facial expressions as Anger emotion (100%). But, as anticipated in Kastendieck et al. [9], a subset of Joy facial expressions exhibits confusion with the Anger emotion (3.6% in our confusion matrix). Our hypothesis attributes this confusion specifically to the Ingratiating Smile, one of the fourteen possible Joy facial expressions.
Figure 11 shows the results of applying LIME to the predictions of two images of Joy emotion labeled as Anger emotion. The facial expressions involved are Eager Smile and Ingratiating Smile, these expressions exhibit the furrowing of the eyebrows and the wide aperture of the eyes. (a) Eager Smile (Joy)
(b) Ingratiating Smile (Joy)
If we analyze the predictions obtained for the Disgust emotion, we realize that it is completely confused with the emotions of Joy and Sadness. Such results are coherent, as the Disgust facial expression shares all their visible facial features with the False Laughter 1 and False Smile facial expressions of Joy, and the Physical Repulsion expression shares all its visible facial features with the facial expressions Crying Closed Mouth and Crying Open Mouthed of the emotion of Sadness (see Table 4). Figure 12 underlines the areas of the eyebrows (relaxed for Disgust and the commented facial expressions of Joy; furrowed for Physical Repulsion and its similar facial expressions of Sadness) and eyes (opened without pressure for Disgust and the commented facial expressions of Joy; closed with pressure for Physical Repulsion and its similar facial expressions of Sadness) of the predictions.
(a) Disgust (b) Physical Repulsion
Concerning the Fear emotion, the results of the prediction are correct in a 75% of the cases as can be seen in the confusion matrix (see Figure 9). However, in a 25% of the cases, the emotion of Fear is labeled as Joy. This could be due to the wide opening of the eyes, as this facial feature is also present in the expressions of Eager and Ingratiating Smile corresponding to the emotion of Joy. But, probably, the incorrect classification is due to one of the limitations stated in Section 2.2 about the texture limitation where a relaxed eyebrow can look like furrowed. In the case of Figure 13c, it seems that the model interprets the shadow of the eyebrow as if it were the eyebrow itself. The images in the Figures 13a and 13b, corresponding to the explanation provided by LIME, show that the areas of interest are precisely located on the eyebrows.
(a) Afraid (b) Terror (c) Worried
The emotion of Joy has an overall recognition rate of 89.3%, which is quite acceptable and is barely confused with Anger (3.6%) (already commented in Section 5.2) and Sadness (7.1%).
Our hypothesis formulated in Section 3 does not expect the confusion between facial expressions associated to the emotions of Joy and Sadness. Figure 14 gives some insight on the superpixels that most influence the confusion of the facials expressions Melancholy Smile and Ingratiating Smile with Sadness emotion. In the case of the image of the Melancholy Smile expression, the highlighted area corresponds to the eyebrow slightly lifted straight up and the inner part of the eye, which appears opened without pressure. The highlighted feature of the eyebrow (slightly lifted straight up) is a facial feature present in four of the six Sadness’ facial expressions, and the only diference resides in the eye opening pressure. This no-detection of a diference in the eye opening pressure could be explained by the inherent limitations present in the geometry of our dataset (see Section 2.2). The case of the image of the Ingratiating Smile expression could be explained, in addition to the diference in the eye opening pressure, by the texture limitation that may cause that a furrowed eyebrow looks like an slightly lifted straight up eyebrow.
(a) Melancholy Smile (b) Ingratiating Smile
The emotion of Sadness has an overall recognition rate of 29.2%, that means that this emotion is dificult to recognize when wearing a mask. In our experiment, this emotion is highly confused with the emotion of Joy (70.8%). As previously commented in Section 5.5, this possibility is considered as plausible because of the limitations of the geometric density of the characters of the dataset and the definition of the eyebrows as texture (see Section 2.2). These limitations could lead to a dificult distinction in the furrowing of the eyebrows and the pressure of the eyes opening. LIME explanations depicted in Figure 15b and Figure 15c seem to confirm this extend as these areas are highlighted as the superpixels of interest. Figure 15a depicts a facial expression of Sadness correctly labelled, where it can be observed that the highlighted areas are inline with the highlighted areas in Figure 15b and Figure 15c. (a) Nearly Crying (b) Miserable (c) Crying Open Mouthed Figure 15: Results of applying LIME to images of Sadness emotion correctly labeled as the emotion of Sadness: (a) Nearly Crying facial expression; and incorrectly labeled as the emotion of Joy: (b) Miserable facial expression and (c) Crying Open Mouthed facial expression.
The Surprise emotion, as predicted, is mostly confused with the Fear emotion (50%). There is also a non-expected confusion with the Joy emotion (25%). This confusion can be explained in the same terms as stated in section 5.4 and in consonance with the LIME explanation (see Figure 16). (a) Surprise (b) Surprise
Based on the confusion patterns observed across all emotions, we can confirm that the hypothesis proposed at the beginning of this study (see Section 3) holds in most cases. Specifically, facial expressions that share the same visible features in the upper part of the face tend to be misclassified as one another when the lower part is masked. This expected confusion was particularly evident among expressions associated with Joy, Disgust, and Neutral, which frequently overlapped with others from diferent emotional categories. However, a few unexpected results also emerged (see Table 11). These outliers may be influenced by subtle geometric and textural limitations in the avatar representation. Overall, the alignment between the predicted and observed confusion patterns reinforces the validity of our hypothesis, while the divergences point to directions for future investigation in masked facial expression recognition.
With the ongoing COVID-19 pandemic and the widespread use of face masks for protection, the dificulty of understanding people’s emotions has become a prominent issue. Empirical evidence and previous studies have unequivocally demonstrated that wearing masks can impair the recognition of facial expressions. The loss of crucial facial information, particularly the concealment of the mouth and chin regions, poses a substantial obstacle to accurately discerning emotions. This loss of information can lead to confusion when trying to recognize an emotion. Such confusions are particularly evident in cases where facial expressions of diferent emotions share visible features in the upper part of the face, such as the eyes and eyebrows.
In this work we have used the theoretical framework proposed by Faigin, where he specifies the facial features that describe each facial expression, to give some insight in the predictable confusions between emotions when the lower part of the face is occluded. We have used explainable artificial intelligence techniques to verify that the areas considered for the prediction of an emotion correspond to the facial features that can cause a confusion. We have not only confirmed that some emotions, such as Anger and Joy, are reliably recognized, while others, like Disgust, Sadness and Neutral emotions, consistently lead to confusion but we have also determined the main factors leading to confusion by analyzing the facial characteristics that influence each of the expressions.
These findings are particularly relevant for the design of interactive systems intended for use in contexts where mask-wearing is prevalent, such as healthcare or public service environments. Incorporating knowledge about likely emotion confusions can inform more robust emotion-aware interfaces, improve adaptive responses from interactive agents, and support the development of compensatory mechanisms (e.g., multimodal emotion detection or user feedback loops) to enhance user experience and communication efectiveness.
This work is part of the Project PID2022-136779OB-C32 (PLEISAR) funded by MICIU/AEI/10.13039/501100011033/ and FEDER, EU. The authors thank the University of the Balearic Islands and the Department of Mathematics and Computer Science for their support.
The authors have not employed any Generative AI tools. [6] F. Grundmann, K. Epstude, S. Scheibe, Face masks reduce emotion-recognition accuracy and perceived closeness, Plos one 16 (2021) e0249792. doi:10.1371/journal.pone.0249792. [7] M. Marini, A. Ansani, F. Paglieri, F. Caruana, M. Viola, The impact of facemasks on emotion recognition, trust attribution and re-identification, Scientific Reports 11 (2021) 1–14. doi: 10.1038/ s41598-021-84806-5. [8] M. A. Pavlova, A. A. Sokolov, Reading covered faces, Cerebral Cortex 32 (2021) 249–265. URL: https://doi.org/10.1093/cercor/bhab311. doi:10.1093/cercor/bhab311. [9] T. Kastendieck, N. Dippel, J. Asbrand, U. Hess, Influence of child and adult faces with face masks on emotion perception and facial mimicry, Scientific Reports 13 (2023) 14848. doi: 10.1038/ s41598-023-40007-w. [10] M. Ventura, A. Palmisano, F. Innamorato, G. . Tedesco, V. Manippa, A. Cafò, D. Rivolta, Face memory and facial expression recognition are both afected by wearing disposable surgical face masks, Cognitive Processing 24 (2023) 43–57. doi:10.1007/s10339-022-01112-2. [11] S. Gil, L. Le Bigot, Emotional face recognition when a colored mask is worn: a cross-sectional study, Scientific Reports 13 (2023) 174. doi: 10.1038/s41598-022-27049-2. [12] T. Saito, K. Motoki, Y. Takano, Cultural diferences in recognizing emotions of masked faces,
Emotion 23 (2023) 1648. [13] G. Carreto Picón, M. F. Roig-Maimó, M. Mascaró Oliver, E. Amengual Alcover, R. Mas-Sansó, Do machines better understand synthetic facial expressions than people?, in: Proceedings of the XXII International Conference on Human Computer Interaction, Interacción ’22, Association for Computing Machinery, New York, NY, USA, 2022, pp. 1–5. doi:10.1145/3549865.3549908. [14] M. F. Roig-Maimó, M. Mascaró Oliver, E. Amengual Alcover, R. Mas-Sansó, Sobre el reconocimiento de emociones y la precisión de los clasificadores, Revista de la Asociación Interacción Persona Ordenador (AIPO) 3 (2022) 55–66. [15] G. Faigin, The artist’s complete guide to facial expression, Watson-Guptill, New York, 2012. [16] M. Mascaró-Oliver, R. Mas-Sansó, E. Amengual-Alcover, M. F. Roig-Maimó, UIBVFED-Mask: a dataset for comparing facial expressions with and without face masks, Data 8 (2023). doi:10.3390/ data8010017. [17] M. Mascaró Oliver, E. Amengual Alcover, UIBVFED: virtual facial expression dataset, PLOS ONE 15 (2020) 1–10. doi:10.1371/journal.pone.0231266. [18] P. Ekman, W. V. Friesen, Facial Action Coding System, Environmental Psychology & Nonverbal
Behavior (1978). [19] L. Colbois, T. d. Freitas Pereira, S. Marcel, On the use of automatically generated synthetic image datasets for benchmarking face recognition, in: 2021 IEEE International Joint Conference on Biometrics (IJCB), IEEE, Shenzhen, China, 2021, pp. 1–8. doi:10.1109/IJCB52358.2021.9484363. [20] J. del Aguila, L. M. González-Gualda, M. A. Játiva, P. Fernández-Sotos, A. Fernández-Caballero, A. S.
García, How interpersonal distance between avatar and human influences facial afect recognition in immersive virtual reality, Frontiers in Psychology 12 (2021). doi:10.3389/fpsyg.2021.675515. [21] M. Marini, A. Ansani, F. Paglieri, F. Caruana, M. Viola, The impact of facemasks on emotion recognition, trust attribution and re-identification, Scientific Reports 11 (2021) 5577. doi: 10.1038/ s41598-021-84806-5. [22] M. T. Ribeiro, S. Singh, C. Guestrin, “Why should I trust you?”: explaining the predictions of any classifier, in: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD ’16, Association for Computing Machinery, New York, NY, USA, 2016, p. 1135–1144. doi:10.1145/2939672.2939778. [23] P. R. Magesh, R. D. Myloth, R. J. Tom, An explainable machine learning model for early detection of parkinson’s disease using LIME on DaTSCAN imagery, Computers in Biology and Medicine 126 (2020) 104041. doi:10.1016/j.compbiomed.2020.104041. [24] S. Sahay, N. Omare, K. K. Shukla, An approach to identify captioning keywords in an image using LIME, in: 2021 International Conference on Computing, Communication, and Intelligent Systems (ICCCIS), 2021, pp. 648–651. doi:10.1109/ICCCIS51004.2021.9397159.