Human emotion recognition based on the weighted integration method using image sequences and acoustic features

People generally perceive other people’s emotions based on speech and facial expressions, so it can be helpful to use speech signals and facial images simultaneously. However, because the characteristics of speech and image data are different, combining the two inputs is still a challenging issue in the area of emotion-recognition research. In this paper, we propose a method to recognize emotions by synchronizing speech signals and image sequences. We design three deep networks. One of the networks is trained using image sequences, which focus on facial expression changes. Facial landmarks are also input to another network to reflect facial motion. The speech signals are first converted to acoustic features, which are used for the input of the other network, synchronizing the image sequence. These three networks are combined using a novel integration method to boost the performance of emotion recognition. A test comparing accuracy is conducted to verify the proposed method. The results demonstrated that the proposed method exhibits more accurate performance than previous studies.


Introduction
humans and computers is actively changing into a bidirectional interface, and a better understanding of human emotions is needed, which could improve human-machine interaction systems [4]. In signal processing, emotion recognition has become an attractive research topic [45]. Therefore, the goal of this human interface is to extract and recognize the emotional state of individuals accurately and to provide personalized media according to a user's emotional state.
Emotion refers to a conscious mental reaction subjectively experienced as strong feeling typically accompanied by physiological and behavioral changes in the body [3]. To recognize a user's emotional state, several studies have applied different forms of input, such as speech, facial expression, video, text, and others [11,13,15,25,39,42,47]. Among the methods using these inputs, facial emotion recognition (FER) has been gaining substantial attention over the past decades. Conventional FER approaches generally have three main steps: 1) detecting a facial region from an input image, 2) extracting facial features, and 3) recognizing emotions. In conventional methods, it is most important to extract appropriate emotional features from the face image. The facial action coding system encodes the movements of specific facial muscles called action units, which reflect distinct momentary changes in facial appearance [8].
In contrast, deep-learning-based FER approaches reduce the dependence between recognition models and preprocessing techniques, such as feature extraction methods, by enabling "end-to-end" learning from outputs to input images. The convolutional neural network (CNN) is the most popular model among several deep-learning models. It convolves input images through many filters and automatically produces a feature map. The feature map is combined with fully connected layers, and the emotional expression is recognized as belonging to a particular class-based output [21]. Recently, various studies have combined facial features and the deep-learning-based model to boost the performance of facial expression recognition [24,38,46].
The speech signal is one of the most natural media of human communication. It contains implicit paralinguistic information and linguistic content, including emotion, about the speaker. Several studies have reported that prosodic features, acoustic features, and voice-quality features imply comparatively abundant emotional significance [28]. The most important issue in the speech-emotion recognition system is the effective parallel use of the extraction of proper speech-signal features and an appropriate classification engine. These features include pitch, formant, and energy features [23,33,41]. In addition, the mel-frequency cepstrum coefficients (MFCC) feature is representatively used in many studies for speech-emotion recognition [26,37,39]. However, because explicit and deterministic mapping between the emotional state and audio features does not exist, speech-based emotion recognition still has a lower rate of recognition than other emotion-recognition methods, such as facial recognition. Therefore, combining appropriate audio features in speech-emotion recognition is critical.
Generally, people recognize the emotions of other people using speech and facial expressions, such as happiness, sadness, anger, and neutrality. According to previous studies, verbal components convey one-third of human communication, and nonverbal components convey two-thirds [19,29]. Facial expressions represent an example of nonverbal components. In terms of perceptual and cognitive sciences, when a computer infers human emotions, it is natural that using speech signals and the facial images simultaneously can be helpful for accurate and natural recognition. However, because the characteristics of the methods to recognize emotions from speech signals and image sequences are different, combining the two inputs is still a challenging issue in the area of emotion-recognition research.
In this paper, we propose a method to recognize emotions by synchronizing speech signals and image sequences. To do this, we design three deep networks. One of the networks is trained using image sequences, which focuses on facial expression changes. Moreover, facial landmarks are input into another network to reflect facial motion. The speech signals are first converted to acoustic features, which are used for the input of the other network, synchronizing the image sequence. Furthermore, we present a novel method to integrate the models, which performs better than other integrated methods. A test comparing accuracy is conducted to verify the proposed method. The results demonstrated that the proposed method shows better performance than previous studies. Therefore, our main contributions in this paper are summarized as follows: & Two deep network models recognize emotions from images, and one deep network model recognizes emotions from speech to reflect temporal representations from two kinds of sequential data. & A method is proposed to learn and classify two different types of data, images and speech, from video data by synchronizing them. & We present a weighted integration method for these three networks with different characteristics, and performance improvement is achieved in terms of accuracy. This paper is organized as follows. Section 2 introduces researches on existing emotion recognition. Section 3 explains the proposed emotion recognition method. Section 4 presents the experiment description and results, and then concludes with Section 5.

Facial emotion recognition
Research on FER has been gaining much attention over the past decades with the rapid development of artificial intelligence techniques. For FER systems, several feature-based methods have been studied. These approaches detect a facial region from an image and extract geometric or appearance features from the region. The geometric features generally include the relationship between facial components. Facial landmark points are representative examples of geometric features [2,30,31]. The global facial region features or different types of information on facial regions are extracted as appearance features [20,36]. The global futures generally include principal component analysis, a local binary pattern histogram, and others. Several of the studies divided the facial region into specific local regions and extracted region specific appearance features [6,9]. Among these local regions, the important regions are first determined, which results in an improvement in recognition accuracy. In recent decades, with the extensive development of deep-learning algorithms, the CNN and recurrent neural network (RNN) have been applied to the various fields of computer vision. Particularly, the CNN has achieved great results in various studies, such as face recognition, object recognition, and FER [10,16,44]. Although the deep-learning-based methods have achieved better results than conventional methods, micro-expressions, temporal variations of expressions, and other issues remain challenging [21].

Audio emotion recognition
Speech signals are some of the most natural media of human communication, and they have the merit of real-time simple measurement. Speech signals contain linguistic content and implicit paralinguistic information, including emotion, about speakers. In contrast to FER, most speech-emotion recognition methods extract acoustic features because end-to-end learning (i.e., one-dimensional CNNs) cannot extract effective features automatically compared to acoustic features. Therefore, combining appropriate audio features is key. Many studies have demonstrated the correlation between emotional voices and acoustic features [1,5,14,18,27,32,34]. However, because explicit and deterministic mapping between the emotional state and audio features does not exist, speech-based emotion recognition has a lower rate of recognition than other emotionrecognition methods, such as facial recognition. For this reason, finding the optimal feature set is a critical task in speech-emotion recognition.

Multimodal emotion recognition
Using speech signals and facial images can be helpful for accurate and natural recognition when a computer infers human emotions. To do this, the emotion information must be combined appropriately to various degrees. Most multimodal studies focus on three strategies: feature combination, decision fusion, and model concatenation. To combine multiple inputs, deep-learning technology, which is applied to various fields, can play a key role [7,22]. To combine the models with different inputs, model concatenation is simple to use. Models inputting different types of data output each encoded tensor. The tensors of each model can be connected using the concatenate function. Yaxiong et al. converted speech signals into melspectrogram images for a 2D CNN to accept the image as input. In addition, they input the facial expression image into a 3D CNN. After concatenating the two networks, they employed a deep belief network for the highly nonlinear fusion of multimodal emotion features [28]. Decision fusion aims to process the category yielded by each model and leverage the specific criteria to re-distinguish. To do this, the softmax functions of the different types of networks are fused by calculating the dot product using weights where the summation of the weights is 1. Xusheng et al. proposed a bimodal fusion algorithm to realize speech-emotion recognition, where both facial expressions and speech information are optimally fused. They leveraged the MFCC to convert speech signals into features and combined the CNN and RNN models. They used the weighted-decision fusion method to fuse facial expressions and speech signals [40]. Jung et al. used two types of deep networks-the deep temporal appearance network and the deep temporal geometry network-to reflect not only temporal facial features but also temporal geometry features [17]. To improve the performance of their model, they presented the joint fine-tuning method integrating these two networks with different characteristics by adding the last layers of the fully connected layer of the networks after pre-training the networks. Because these methods mostly use shallow fusion, a more complete fusion model must be designed [28].

Preprocessing
When constructing a video emotion database, the actors start and finish expressing emotions according to the instructions of the experimenter. Therefore, as shown in Fig. 1, the database was often divided into three sections-the section for the actor to express emotions, the section to prepare the emotional state, and the section finishing expressing emotions. For this reason, the need to determine whether a given speech signal and image sequence should be classified as the acting section or the silence section arises in many emotion-recognition systems. When nonspeech sections are included in the learning or testing process, they can provide unnecessary information and become an obstacle. For more accurate processes, this section describes removing these nonspeech sections. Because the signal energy value of the speech-signal segment is larger than that of the nonspeech-signal segment, an absolute integral value (IAV) reflecting the energy value was used. The IAV value was computed using Eq. (1): where X is the recorded signal,Δtis the time interval, N is the number of samples, and i is the sample index. First, the IAV feature vector must be extracted from the interval of the signal. Then, it is imperative to calculate the maximum and minimum values and determine the threshold value with a 10% difference between these two values. An example of determining the threshold is shown in Fig. 2.
The process of selecting the start point for a speech interval includes a point at which the window is larger than the IAV value. If the extracted IAV value was smaller than the IAV threshold, the endpoint was determined. The points were quantized using Eqs. (2) and (3) so that the speech signals and image sequences were synchronized.
To map 30 Hz (33.33 ms) of the sampling rate of the image sequence, the window size of the speech signals was 1600 (33.33 ms). Accordingly, the input of an image sequence and speech signal at a point used one image and 1600 speech-signal data, respectively.

Image-based model
To recognize emotions from a facial image sequence, we used two deep-learning networks. The first network captures temporal changes in appearance by combining the CNN and LSTM models. The proposed CNN and LSTM models are illustrated in Fig. 3.
In general, the length of image sequences varies in every video, but the input length of a deep network is usually fixed. Therefore, the length of the image sequence must be fixed. In this study, we set a time step of the image sequence to ten. The network infers an emotion every 0.3 s. Before inputting an image sequence to the network, all images were converted to grayscale. Then, the faces in the input images were detected, cropped, and rescaled to 64 × 64. The common 2D-CNN layer used still images as input. We combined CNN layers and LSTM layers to deal with image sequences.
The CNN layers of this network used the image sequences as input without sharing weights along the time axis. Thus, the filters played different roles depending on the time. Each image Fig. 2 An example of determining the threshold along the time axis was converted to feature maps through each convolutional and pooling layer. After convolving the images, all output passed through rectified linear unit activation functions. The feature maps were stacked in time order so that they were input into the LSTM layers. The output of the LSTM layer was connected with the fully connected layers, and the last layer inferred the probability of each emotion through the softmax function. To train the whole network, the AdaDelta optimizer method was used, and the weight-decay and dropout methods were used for regularization.
The network that input the landmarks was derived from a previous study. Because landmarks generally reflect facial motion, they complement another model to infer facial expression. First, landmarks can be considered to be a 1D vector as follows: Where n is the total number of landmark points at frame t, andX (t) is a2 × ndimensional vector at t. In addition,x  i is the standard deviation of the x-coordinates at frame t. This process is also applied toy t ð Þ i . We concatenated the normalized vector along the time step. The vector is used as input to the network, as shown in Fig. 4.
The network receives the normalized vector as input, and the last layer infers the probability of each emotion through the softmax function. The dropout methods are used between each fully connected layer for regularization.

Speech-based model
Because verbal components convey one-third of human communication, it is natural that using speech signals and a facial image simultaneously can be helpful for accurate and natural recognition. Therefore, we propose a reasonable feature combination that can improve emotion-recognition performance using an RNN, complementing the FER. In previous emotion-recognition combining speech signals and image sequences, many studies used only the MFCC feature or images converted from a mel-spectrogram [12,28,40]. We surveyed acoustic features used for many speech-emotion recognition studies and composed an optimal feature set by analyzing and combining the interconnectivity of each feature. Harmonic features reflecting the harmony of speech are used, which were less used for previous studies. First, we selected specialized features for emotion recognition through individual analysis and found the optimal feature set by recombining features.
In total, 43 features were extracted and are used in this paper: 11 spectral-domain: spectral centroid, spectral bandwidth, 7 spectral contrasts, spectral flatness, and spectral roll-off; & 12 chroma: 12 dimensional chroma vectors; and. 7 harmonic features: inharmonicity, 3 tristimuli, harmonic energy, noise energy, and noisiness. If the range of each attribute value of the learning data is greatly different, the learning will not work efficiently. For example, if the range of a feature vector A is 1 to 1000, the range of another feature vector B is 1 to 10, and the value of A is larger, it seems as if it has a significant effect on the neural network while B seems as if it does not relatively affect the network. Thus, transforming each property value into the same range is necessary before the learning process, and this process is referred to as "feature scaling." In this study, we normalized the features using the standard-score method, which considers the range and variation of the values. The equation of this scaling method is as follows (2).
where x ' is a normalized vector, x is an input vector,xis the average of x, andσis the standard deviation of x.
After windowing the speech signals, the signals are converted to acoustic features, and the features are input into the LSTM layers. The output of the LSTM layer is connected with the fully connected layers, and the last layer infers the probability of each emotion through the softmax function. The whole speech-based model is illustrated in Fig. 5. The weight-decay and dropout methods are used for regularization.

Weighted joint fine-tuning
The previous study [17] proposed a joint fine-tuning method that integrates two networks. After pretraining the networks, the networks were reused. They integrated the two networks by adding the last layers of the fully connected layer of the networks. Then, the linear fully connected networks were retrained, which achieved better results. In this paper, we designed an integration method that weighted each model in the integration process. The last layers were integrated using Eq. (7): Based on the preliminary experiments, we set W 1 , W 2 , and W 3 to 0.2, 0.2, and 0.6, respectively. Each model was trained using softmax, and pretrained models were integrated using Eq. (7). Finally, the integrated model calculates the probabilities for emotions using another softmax function.

Ryerson audio-visual database of emotional speech and song dataset
The Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS) is a video database of emotional speech and songs in North American English, classified into eight emotions as shown in Fig. 6, including neutral, calm, happy, sad, angry, fearful, disgusted, and surprised. The database comprises information from 24 professional actors, and each actor has 60 audio-visual (AV) items and 44 song items, for a total of 104 data points. Each recorded production of an actor was available in three modality formats: AV, video only, and audio only. Among these, we used 24 × 60 × 3 = 4320 AV data. For validation of this database, 247 raters each rated a subset of the 7356 files. For reliability, a further 72 raters provided intra-rater test-retest data. Validation was achieved by asking the raters to label the expressed emotion. In RAVDESS, contrary to traditional validation methods, for facial recognition databases, accuracy, intensity, and genuineness must be verified for emotion measurement of all presented stimuli because orofacial movement, where movements are tied to the lexical content, interacts with movements related to emotional expression. To select the appropriate stimuli, the "goodness" score was imposed. The goodness scores ranged between 0 and 10 and are a weighted sum of the mean accuracy, intensity, and genuineness measures. The equation was defined such that stimuli receiving higher measures of accuracy, intensity, and genuineness were assigned higher goodness scores.

Baselines
This section describes the baseline algorithms for model integration.

Multi-input model
When using different input types for classification, the models should be designed so that each model reflects the characteristics of the data. To combine the models with different inputs, a layer that can connect tensors was used. Models inputting different types of data output each encoded tensor. The tensors of each model can be connected using the concatenate function. Then, the final outputs of each model were integrated by adding the softmax function. Fig. 7 describes the example of the multi-input model.

Feature concatenation
To recognize human emotions by learning both facial and speech data, facial data were converted to a feature map by inputting the data into the CNN. Then, we merged the feature map with features from the speech data, as shown in Eq. (8).
Where f is the feature map from the facial data, and s is the feature of the speech data. Lastly, emotions were classified using the feature vector x as the input of the LSTM model in a timeordered sequence.

Joint fine-tuning
To incorporate models with different data, first, each model with different data was trained.
Only fully connected layers in the pre-softmax classification stage from already trained models were used as new integrated models. The weight values from the already trained models were frozen, and the fully connected layers from each model were retrained. Then, the integrated Fig. 7 Examples of the multi-input model model classified emotions using another softmax. The softmax function for training in each model was used when calculating the loss function, and only the softmax function of the integrated model was used when predicting.

Results
As mentioned in Section 4.1, this study used the AV data in the RAVDESS dataset to test performance. The dataset comprised eight emotions, and we used only emotional speech, except the emotional song data. In the RAVDESS, all sequences start and end with a silence section, which was removed through preprocessing (Section 3.1). In addition, image and speech data were separated by synchronizing with each other. The data of a training set were as follows: image data (10, 64, 64), landmark data (980), and speech data (10,43). To verify the performance of the proposed method, ten-fold cross-validation was performed (Table 1). When learning the model proposed in this paper using image and speech data with the joint fine-tuning method, the accuracy was 86.06%. When learning the model using the multi-input model and feature combination, the accuracy was 81.93% and 78%, respectively. The model learned using weighted joint fine-tuning demonstrated the greatest result at 87.11%. This also increased the accuracy by about 2.5% compared to the model using only image data (84.69%).
Jung et al. proposed the model recognizing facial expressions using image data, constructing two small deep networks that complement each other [17]. The model proposed by Jung et al. exhibited an accuracy of 85.72% in the RAVDESS dataset. Wang et al. [40], Ma et al. [28], and Hossain et al. [12] proposed models integrating a CNN model input with image data with a 2D CNN model input with speech data by converting the speech signals to melspectrogram or spectrogram images. The studies, which converted the speech data to a spectrogram to integrate the image and speech, demonstrated an accuracy of about 75% to 77%. The proposed model integrating image and speech data using acoustic features produced a greater result by about 10% than the other integration methods. The multiple-input model integrating each model using the concatenate function is simple to use, but it may not maximize the ability of the networks. We fine-tuned the softmax functions of the pre-trained networks, considering the characteristic of each input, to maximize the ability of the networks. For this reason, the proposed method can produce more accurate results than the multiple-input model. Lastly, most of the previous studies using the RAVDESS dataset used only speech data by converting the speech data to an acoustic feature. They exhibited an accuracy of 64.17% and 74%, respectively. Thus, the proposed model dramatically increased the accuracy (87.11%) by integrating the image and speech data.

Conclusions
We presented three networks to reflect the characteristics of each input data. One of the networks was trained using image sequences, which focus on facial expression changes. In addition, facial landmarks were input into another network to reflect facial motion. The other network used acoustic features from speech data as input. These three networks were combined using a novel integration method to boost the performance of emotion recognition. To investigate the performance of our model, we tested the recognition accuracy with previous studies on the RAVDESS dataset. According to the results, our model achieved the best recognition rate against facial and speech-based studies. Furthermore, we demonstrated that our weighted joint fine-tuning method exhibited better performance than other methods.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.