Facial Expression Recognition from a Single Face Image Based on Deep Learning and Broad Learning

With advances in computer vision and arti ﬁ cial intelligence technology, facial expression recognition research has become a prominent topic. Current research is grappling with how to enable computers to fully understand expression features and improve recognition rates. Most single face image datasets are based on the psychological classi ﬁ cation of the six basic human expressions used for network training. By outlining the problem of facial recognition by comparing traditional methods, deep learning, and broad learning techniques, this review highlights the remaining challenges and future directions of deep learning and broad learning research. The deep learning method has made it easier and more e ﬀ ective to extract expression features and improve facial expression recognition accuracy by end-to-end feature learning, but there are still many di ﬃ culties in robustness and real-time performance. The broad learning system (BLS) is a broad network structure that is expanded by increasing the number of feature nodes and enhancement nodes appropriately to reinforce the structure and is also e ﬀ ective in facial expression recognition. However, outliers and noises in unbalanced datasets need BLS to solve in the future. Finally, we present several problems that still need to be addressed in facial expression recognition.


Introduction
Facial expression is an important way for humans to present their emotions. They can affect our daily lives by changing our attention, perception, and memory, helping us understand the intentions of others. As an information carrier, facial expressions can accurately express the true emotions of human beings. Humans can also learn the inner thoughts of others through facial expressions [1,2] [3]. According to a study by psychologists Mehrabian et al., facial expressions are prominent in daily human communication, with a transmission rate of 55%, far more significant than speech (38%), language (7%), and so on [4]. Face expressions, on the other hand, follow the rules of facial muscle movement and are unaffected by gender, age, race, or cultural background [5]. As a result, facial expressions are a powerful tool for detecting emotions. Facial expression recognition research is crucial for advancing artificial intelligence and other domains. The research on facial expression recognition based on computer technology can enable intelligent devices such as robots to better understand and identify human emotions, so as to actively judge human emotions, better serve humans, and achieve barrier-free interaction between humans and machines [1,6,7].
The performance of deep convolutional learning technology in natural language processing, computer vision, and other information technology sectors has grown increasingly significant due to ongoing in-depth research on deep convolutional networks. People's perceptions of deep learning convolutional neural networks have shifted as a result [8]. The application and performance of deep learning technology in image classification and recognition are better than the application and performance of previous algorithms in terms of recognition accuracy, and recognition speed as more universities and scientific research institutions conduct more in-depth research. As a result, more researchers are becoming aware of deep learning technologies. The deep learning technique eliminates numerous artificial technological ways to extract features and instead performs end-to-end feature learning directly on the picture.
As a result, we use deep learning-based technology in facial expression recognition to help it extract more effective and more prosperous expression characteristics, considerably enhancing facial expression identification accuracy. Face expression recognition is prone to complex issues such as difficulties extracting facial characteristics, low recognition accuracy, and slow recognition speed [9].
Although the facial expression recognition system has a wide range of applications and a bright future, several technical issues with the actual product landing still have to be resolved. In the real world, numerous adverse external effects alter the picture of the face, failing face frame extraction or poor recognition degree, and facial expression recognition in the video image must have a quick recognition rate to be successful system assistance. More facial expression recognition applications will gain traction due to their high accuracy and speed.
Deep learning research in facial expression has increasingly risen as deep learning technology in the field of image categorization has advanced. As a result, we focus on deep learning approaches in the context of expression recognition research. First, we will go through the history of face recognition technology and introduce some of the most popular expression datasets. The impacts of traditional and deep learning technology on different data sets are compared, and facial expression recognition technology based on deep learning is thoroughly examined from the perspectives of facial recognition challenges and expression recognition algorithms. Furthermore, the challenges and future directions in deep learning-based expression recognition are discussed.

Overview of Facial Expression Recognition Technology.
As indicated in Figure 1, surprise (Su), fear (Fe), disgust (Di), happy (Ha), sadness (Sa), and anger (An) are the basic human emotions. In the past, automatic face expression interpretation from picture sequences was primarily a research area for psychologists. However, Suwa [10] did a preliminary investigation in 1978. Even though many academics feel the six basic emotions are culturally unique and not universal, the categorization approach based on the six fundamental expressions has been widely accepted by researchers and has aided the growth of expression recognition [11].
Facial expression recognition (FER) is a method that consists of three steps: preprocessing of facial images, extraction and representation of facial expression features, and recognition of facial expressions [13]. Figure 2 depicts the fundamental procedure. Facial image preprocessing is a stage of video processing that automatically recognizes faces in an input picture or sequence, detects faces for each frame, just detects faces in the first frame, and then tracks faces for the rest of the video sequence. Face analysis is complicated by variations in facial appearance caused by changes in position and illumination. As a result, preprocessing is necessary before analysis to align and normalize the visual-semantic information communicated by faces. Following the prepro-cessing of face images, the next step is to extract usable data or information from the images, and the feature extraction procedure is critical for FER. This technique allows for a smaller and richer collection of characteristics, such as face edges, corners, diagonals, and other critical information (such as the distance between lips and eyes and the distance between two eyes), allowing the model to learn from the training data more rapidly. The FER pictures are categorized after face detection, preprocessing, and feature extraction phases.

Facial Expression Database.
Most expression recognition methods must be tested on a set of expression data, and the generation of expression data is thus a major driving factor in the advancement of expression recognition technology. Traditional expression recognition algorithms are sparse learning nonnegative matrix factorization (NMF) [15], local binary patterns (LBP) [16], LBP on two orthogonal planes (LBP-TOP) [17], and so on. Since 2013, however, general emotion recognition competitions such as FER2013 [18] and EmotiW [19][20][21] have gathered large-scale facial expression image data, primarily from the Internet or movie clips, that align with the real scene and effectively communicate promote the facial expression recognition transitions from laboratory controls to real-world scenarios. In image acquisition, facial expression images are usually acquired under laboratory conditions and subjected to specific preprocessing. At present, the well-known datasets include Japanese Female Facial Expression (JAFFE) and Fer2013. The main information of each dataset is shown in the following Table 1.
The CK+ (The Extended Cohn Kanade) expression database is a 593 video sequences laboratory-controlled database. The CK+ expression database video sequences were gathered from 123 participants, more than 60% of whom were women, all of whom were between 18 and 30 years old [22]. The following are the features of the video sequence. The lowest frame duration is ten frames, while the maximum frame duration is 60 frames. The face begins in a neutral state, with no expression, and gradually progresses to a peak of emotion.
The JAFFE (Japanese Female Facial Expression) database contains 219 images of 10 volunteers' facial expressions [23]. Six fundamental emotions were gathered from all volunteers, three to four photographs for each category and a neutral face image. The photos of facial expressions are stored in the database as black and white photographs that have been preprocessed.
There are 2,880 video sequences in the Oulu-CASIAP1 expression database from 80 volunteers [24]. The video series begins with a neutral frame and progressively evolves to the peak emotion, similar to the CK+ facial expression database. FER experimental verification typically only employs 480 video sequences acquired by the Visual Identity System (VIS) in standard indoor lighting, with each sequence beginning with a frame (neutral face) and ending with three peaks (expressive face) as experimental data.
The expression database RAF-DB (Real-world Affective Faces Database) is based on real-world scenarios, and the expression images in it are all retrieved from an Internet search engine in the form of emotion-related and age, race, and gender-related keyword combinations, totaling 29,672 pieces [25]. Each emoticon image was formed from roughly 40 persons, and the labels were annotated online by 315 annotators (students and university personnel) who underwent an hour-long online instruction on the psychological understanding of emotions.
The FERP plusM facial expression dataset grew out of the FER2013 facial expression dataset [26]. All pictures were  3 Wireless Communications and Mobile Computing scaled to 48x48 pixels after being preprocessed. The dataset is commonly separated into 28,709 photos for training, 3,589 images for validation, and 3,589 images for testing when used for experimental validation.
Affected is the world's most extensive library of realworld face expressions, with over 1 million photographs gathered from the web by searching emotion-related categories using various search engines. It offers facial expressions in two emotion models (category and dimensional), with 450,000 manually annotated with seven fundamental expressions (contempt, sorrow, fear, surprise, pleasure, disgust, and rage) and neutral expressions [27].
The Acted Facial Expressions in the Wild (AFEW) database is used to create the SFEW (Static Facial Expressions in the Wild) dataset. The expression video clips from movies that are designated as six fundamental expressions (anger, contempt, fear, sorrow, pleasure, and surprise), as well as neutral expressions and facial movements, comprise the material of AFEW. Units (Action Units, AUs) are labeled to mimic real-world face emotions as closely as possible [19].

Difficulties in Facial Expression Recognition Research.
There are still numerous obstacles and problems to be solved when the scientific focus of expression recognition work switches to challenging real-world unconstrained environmental settings.
(1) Lighting Problem. It is possible to lose vital information in specific face areas if there is insufficient or no light and the light is too bright. There is a loss of crucial knowledge. As a result, lighting conditions are an issue for expression recognition, and they can significantly impact the algorithm's robustness (2) Lack of High-Quality Public Datasets. Data bias and inconsistent labeling are common concerns when utilizing diverse datasets due to the diversity and subjectivity of the distribution of expression categories (3) Occlusion and Pose Changes. The FER model cannot get adequate facial information because the face is obstructed or the posture rotation is too great, and when the obtained facial information is less than 60%, the expression recognition cannot be conducted typically (4) Multimodal Recognition. Although facial expression recognition may produce good results from visible face photographs on its own, merging it with additional models into an integrated system, such as adding a voice model, adds more information and helps to enhance model dependability (5) Big Data Problem. Data is measured in terabytes in many scientific and commercial applications, posing issues for data storage, transport, and processing in FER systems (6) Privacy Protection. Many expression recognition algorithms rely on high-resolution face photos of users, yet many researchers pay little or no attention to ensuring that users' visual privacy is protected

Expression Recognition Algorithm
The science of expression recognition has advanced significantly over the years. Existing approaches may be split into standard machine learning algorithms and deep learning algorithms. Figure 3 depicts the framework for the two sorts of approaches. Face preprocessing, feature extraction, and classifier prediction are three discrete manual stages in traditional expression recognition methods, as depicted in Figure 3(a). As demonstrated in Figure 3(b), deep learning networks combine feature learning and classification into a single framework, eliminating the need for complicated manual methods.

Traditional Expression Recognition Algorithms.
Preprocessing is the initial stage in classic expression recognition systems since low-quality photographs influence the system's accuracy. Preprocessing can help you eliminate as much task-irrelevant data as feasible while boosting task-relevant data. Face detection, histogram equalization, picture normalization, and face alignment are typical image preprocessing techniques. To extract the critical information that best recognizes expressions based on the digital signal of the image, well-designed feature extractors are required. Traditional approaches may be classified into the following kinds based on the extracted characteristics. Traditional facial expression recognition systems mainly require an artificial design to extract facial expression elements in real-life applications. Artificially constructed features frequently fail to adapt to complicated and changing expressions, resulting in an inadequate final expression categorization outcome.

Expression Recognition Algorithm Based on Deep
Learning. Hand-crafted features are time-consuming and inaccurate and are constrained mainly by the specified algorithms, which can only represent information under certain formal constraints. It has progressed from small sample jobs in typical laboratory situations to large-scale sample tasks in real-life scenarios in facial expression recognition. Deep learning has been applied in numerous computer vision tasks in recent years. Autoencoder and generative adversarial networks (GANs) are also used in several expression recognition methods [5,30].

Expression Recognition Based on Convolutional Neural
Network. CNN-based expression recognition research content is vast, covering architectural design and loss function design [31]. Yu and Zhang [32] built an eight-layer CNN architecture with five convolutional layers and three fully linked layers. Yu et al. preprocess pictures using a welldesigned multilevel face detection framework (containing three current state-of-the-art detectors) before using a CNN architecture for end-to-end feature learning and classification, as illustrated in Figure 4.

Expression Recognition Based on Recurrent Neural
Network. Recurrent neural network (RNN) is a recurrent neural network with a memory function and can learn time series data's dynamic evolution information. The RNN's nodes are linked together in a chain, and the input sequence data is processed recursively in the time evolution direction. RNNs are typically employed for temporal logic applications, such as natural language processing and video comprehension. Many studies use RNN or its variations, such as long short-term memory (LSTM) and bidirectional LSTM, for facial emotion recognition. The feature vector of the picture is commonly obtained as input data using a manual extractor or another deep learning model in these approaches.
As demonstrated in Figure 5, Yu et al. [33] introduced an end-to-end trainable deep learning model called STC-NLSTM for identifying medium expressions in picture sequences. SP3DCNN and NLSTM are two critical components of STC-NLSTM. 3DCNN is at the front end of the

Expression Recognition Based on Deep Belief Network.
Deep belief network (DBN) is a generative model with a restricted Boltzmann machine in the layer structure next to it. By stacking greedy training, data satisfying a given distribution may be generated. DBNs are frequently connected to classifiers or provide superior initialization parameters for deep neural networks that are turned into discriminative models when dealing with classification challenges. DBNs have been used to solve problems in various fields, including facial expression recognition.
Lv et al. [34] used histogram of oriented gradients (HOG) features to train DBN as a facial key region detector. They separated face image blocks into different sizes and extracted different local areas from each block. The best detection findings are employed in deep neural networks to recognize expressions. As illustrated in Figure 6, Kurup et al. [35] created a semisupervised DBN network that comprises two parts: (a) for unsupervised feature learning, a DBN comprised of constrained Boltzmann machines; (b) for supervised feature learning and classification, softmax activations coupled at the top of the DBN. Face identification, picture local patch extraction, manual feature extraction, and feature dimensionality reduction were among the preprocessing steps Kurup et al. took. The expression image is reduced to a low-dimensional representation, which is then fed into the proposed DBN network for feature learning and classification.  Wireless Communications and Mobile Computing Deep learning successfully avoids human interference by performing end-to-end feature learning directly. The use of deep learning technology in facial expression recognition makes it easier and more effective to gather expression characteristics, boosting facial expression recognition accuracy. In 2016, Jeon et al. [36] presented a histogram of oriented gradients (HOG) for face identification, followed by the application of convolutional neural networks for feature extraction and classification tasks, with excellent results in various datasets. In 2017, Al-Shabi et al. [37] introduced a technique for integrating scale invariant feature transform (SIFT) with convolutional neural networks and created a hybrid classifier CNN-SIFT, a collection of two feature extractions with a very excellent effect on small sample detection and identification. The CK+ expression dataset is relatively small. However, the accuracy of this method on CK+ after training achieved 99.4%. According to the findings, deep learning and convolutional neural networks are superior to classical machine learning techniques in image identification, recognition, and classification [38].

Expression Recognition Algorithm Based on Broad
Learning. The broad learning system (BLS) network is different from the general deep learning network [39]. It builds a horizontal network structure, and its network structure is not fixed during the training process. The widening network has reached the optimal classification effect.
Compared with the general deep learning network, the BLS has a simple structure and fewer parameters with horizontal expansion. The basis of the BLS network is the random vector functional-link neural network (RVFLNN) [40]. The network structure of RVFLNN is shown in Figure 7(a). For a given input X, a normal neural network will multiply the input X by a weight and add a bias to the next hidden layer. Nevertheless, the RVFLNN network is more than that. It also multiplies the input X by a set of random weights and adds a bias to get an enhancement layer through nonlinear mapping of the activation function. Finally, the input X is connected to the output layer together with the enhancement layer data. Based on the RVFLNN network, the BLS network has been improved.
On the one hand, BLS does not directly use the original input data as the network input but does a linear transformation, which is equivalent to feature extraction and uses the transformed feature data as the feature layer of the network. The feature layer is nonlinearly mapped to generate an enhancement layer, and finally, the feature layer and the enhancement layer are combined as the network's input layer. BLS has a core incremental learning part, and its network structure is not fixed. During the training process, incremental learning can be achieved by increasing the number of feature nodes or enhancement nodes to improve the network model. A basic BLS network structure is shown in Figure 7(b). The BLS system was tested using the extended Cohn-Kanade dataset (CK+), and the BLS method was very effective in facial expression recognition compared to convolutional neural networks [41].
The general neural network has many network parameters due to deep layers and many neural nodes, and the training is very time-consuming. The BLS network model has a simple structure and few adjustable parameters and is suitable for data with low feature dimensions. When the network training fit is insufficient, the number of nodes in the feature layer or enhancement layer can be appropriately increased to improve the recognition effect, and the training process is simple and fast.

Summary and Outlook
Expression characteristics are complicated and variable, and they account for a significant percentage of emotional computing. Face recognition research using computer technology might help intelligent objects like robots better comprehend and recognize human emotions, allowing for barrier-free contact between humans and machines.  (1) Head Posture Is Causing Self-Occlusion [39]. Occlusion is a prevalent phenomenon in unconstrained natural landscapes and a tough challenge in computer vision. Head posture changes are common in expression photos, causing extreme self-occlusion, which has two immediate effects: losing essential facial information and altering the look of expressions. Local parts such as the eyes, lips, and others are not visible; the same emotion will seem different in different head situations.
(2) Overfitting Problem in Deep Expression Recognition. The model in deep learning generally has a very deep hierarchical structure, and the represented function is quite complicated, which makes overfitting very easy. The overfitting problem is especially obvious when used to expression recognition. Because human emotions are so varied, face pictures are frequently a mixture of numerous expressions rather than a single one. As a result, expression photographs have a natural quasiblur. On the other hand, hard labels can only reflect one category and cannot explain real-life expressions. Current algorithms usually use hard labels, which can easily aggravate model overfitting and severely reduce the generalization ability [40] (3) The Challenge of Extracting Discriminative Expression Features. The key to expression recognition in machine learning is to learn discriminative features from pictures and then do class matching [41]. A high degree of interaction between parameters such as backdrop and light intensity generates the picture  Figure 7: (a) RVFLNN network structure [40]. (b) BLS network structure [42]. 8 Wireless Communications and Mobile Computing in an unconstrained natural scene, while numerous irrelevant qualities are jumbled on the surface. As a result, the image's complexity is tremendous, and there is much noise, making learning expression characteristics incredibly tough.
(4) The Selection and Extraction of Feature Information Is the Basis for Building a Wide Neural Network Model. The video emotional features extracted by the convolutional network can hardly describe the facial expression changes in the video, so it needs to be combined with other texture features to improve the recognition rate of various emotions.
The rapid development of deep learning in recent years is evident to all, and new structures are constantly emerging, which also drives the development of facial expression recognition technology, and the accuracy on many public datasets keeps rising [42]. However, with the gradual deepening of the number of layers, the number of parameters becomes too large, requiring more resources to optimize [43]. On the other hand, broad learning has the advantages of fast computation, incremental online learning, and ease of scaling [35]. However, it is still necessary for BLS to solve the outliers and noises problem [44].The future will focus more on facial expression recognition in-the-wild conditions [5], such as education, medical, traffic, criminal investigation, and business. Some factors such as lower image resolution, occlusion, and pose changes will make the expression discrimination and understanding more challenging [45]. Facial expression recognition that incorporates multimodal information is helpful to realizing more detailed emotional understanding, and it is the future development direction [46,47]. A question worthy of in-depth study is whether the two algorithms can always maintain good computing power and generalization performance.

Data Availability
The data used to support the findings of this study are included within the article.

Conflicts of Interest
The authors declare no potential conflicts of interest for the research, authorship, and/or publication of this article.