Emotion Recognition from Facial Expression Using Machine Vision Approach

Journal of Applied and Emerging Sciences Vol (10), Issue (01) [35] Abstract--In this study emotion-based face expression recognition framework has been proposed using a machine vision (MV) approach. The face emotion dataset has been collected local survey in Bahawalpur city dataset divide into 3 classes happy, sad, and angry. A total of 600 images of size (256 x 256) were transformed into a gray level format and employed a median filter for noise removal. Three nonoverlapping regions of interest (ROIs) of size (50 x 50) have been taken and analyze 1800 (600 x 3) ROIs on the overall dataset. Total 45 Statistical features named as texture, histogram, and binary features were extracted. Select optimize features using the correlationbased feature section technique. The optimized dataset employed of MV classifiers namely random forest (RF), logistic (Lg), and J48 are obtained very promising accuracy 96.33%, 95.67%, and 95.33% respectively.


I. INTRODUCTION
N human communication, facial emotions play a very important role that helps in intention understanding [1]. Generally, people conclude other people's emotional states, such as happy, sad, and angry, using facial expressions. Oral components transmit one-third of human communication and non-verbal components transmit two-thirds [2]. Among the various non-verbal components, with an emotional meaning, facial expressions are one of the main channels of information in interpersonal communication [3]. Facial expression is an important part of nonverbal communication; because the facial gesture has an important part in explaining human actions. To obtain the required evidence, the facial expression data is divided into several sections which are applied in different methods to machine vision. The zone of the nose, lips, and eyes is transformed during the dialog [4].
The researcher proposed automated face analysis using a multi-feature approach. The FASSEG dataset is used for experimentation. Segmentation performs on the bases of nose, eyes, and skin, and random forest, support vector machine (SVM) classifiers used for classification and obtained 96.25% accuracy [5].
The researcher proposed conventional facial emotion recognition (FER) approaches are described along with deep learning (DL) based FER approaches and obtain very promising accuracy result of 72.65% [6]. The researcher proposed an automatic assessment of gender and FER. The idea based on image segmentation in six parts and label it. Random decision forest classifier and SVM classifier used for classification and obtain 96.25% accuracy [7].
The researcher proposed Male and female adolescents with conduct disorder associated with impairments FER. Concurrent eye tracking was used to relate categorization performance to participants' allocation of overt attention [8].
The researcher described the promising sectors in the field of FER, processing images streamed in real-time from a mobile device, which is health care. The proposed approach based on convolutional neural networks (CNN) and achieved better classification results [9].
The researcher described facial expression datasets such as eyes, nose, lips, and chin, etc. The overall accuracy is 90% using K-Nearest Neighbor (KNN) algorithms [10].
The researcher described edge detection algorithms for eyes and lips variation during human communication. canny edge detection (CED) provides better results on following facial expressions such as Sad 91.33%, Smile 99.44%, and Surprise 95.71% [11].

II. IMAGE DATASET
In this study digital image dataset has been collected from the local survey in Bahawalpur using a digital camera with 12 Megapixels. Three types of classes consider during the data collection phase first one is happy, the second is sad and the last one is angry. All the images based on real human emotions. For [36] this study 200 humans consider and divide into 3 classes, so a total 600 (200 x 3) images size (256 x 256) were collected for experimentation. The collected dataset sample is shown in Figure 1. In this study, first, we collected an image dataset for experimentation. All the acquired image datasets standardized on the behave of ground truth. In the second step, color digital images transformed into gray level format [12]. After that using the CVIP tool select non-overlapping region of interest (ROIs) size (30 x 30) and (50 x 50). At the third step, texture-based features extracted named as texture (T), binary (B), and histogram (H) features. A fourth step correlation-based feature selection (CFS) applied for feature selection [13]. At last step machine learning classifiers [14] named as random forest (RF), logistic (Lg), and J48 applied on selected features dataset. All these experiment phases organized using Intel(R) core i7 processor 3.4Giga Hertz (GHz), 4 GB RAM, and 32-bit window 8.1 operating system. Now we describe proposed algorithm then proposed diagram in Figure 1.

Happy Sad Angary
Step 1 to 4 Step 1: Transformed color to gray level and select non over lapping ROIs.

B. Image Preprocessing
Image preprocessing is a very important part of image analysis during image dataset collection some noisy image is the part of the dataset. So, for data standardization, we deal with the preprocessing stage. Firstly, crop all the images and set stander size 254 x 254, and all the image transformed into a gray level format. Some noise removal filters were applied named as a median filter [15], after that selected non-overlapping ROIs.
The ROIs selection is the main part of this study our focus is on eyes, nose, and lips these are the more dominant portion which effected based on human mood. We select multiple ROIs with different size (30 x 30) and (50 x 50).

1) Binary Feature
Binary features recognize the substances and the shape of the entity in image processing based on Area, Second Axis, Euler number, Central Area and Projection [16].

2) Histogram Feature
Histogram shapes have been calculated the information about image on the basis of gray level and number of pixels [17]. The 1 st order histogram W(h) as follows in equation (1).
Here N is total pixel in the image and M(h) is total pixel at gray level of K.
The 1 st order histogram are used following method that are mean, standard deviation, skewness, energy and entropy for statistical calculation. Mean define as follow in equation (2).
Skewness define as follows in equation (4).
Energy is defined in equation (5).
Entropy is defined in equation (6).

3) Texture Feature
Energy evaluated smoothness by calculating the distribution between gray levels [18] are define in equation (7).
Here are the values in the co-occurrence matrix by distribution values of pixel.
Correlations method defines as follows in equation (8).
Here and are the means of a and b respectively.
Entropy is measuring the content information of image. Entropy defines as follow in equation (9).
Inverse difference method is measuring the local homogeneity of image that defines as follow in equation10. [38] Inertia method is measuring the contrast that defines as follow in equation (11).
A total of 45 extracted statistical features (B + H + T) has been examined for each ROIs. It has been observed that the total feature vector space is 81000 (45 x 1800) on the basis of an employed dataset.

D. Classification
Correlation-based feature selection (CFS) approach used for select 9 optimize features for the classification. Three machines vision classifiers named random forest (RF), logistic (Lg), and J48 were employed using a cross-validation (10-fold) approach.

IV. EXPERIMENTS AND RESULTS
In the first step, machine vision classifiers employed of emotion dataset were ROIs size (30 x 30) and observe that impressive classification accuracy is random forest (RF), logistic Lg, and J48 were 95.33%, 94.67%, and 94.33% respectively as shown in Table 1. In the second step for improving classification accuracy result we increase the size of ROIs (50 x50) and observe very promising classification accuracy is RF, Lg, and J48 were 96.33%, 95.67%, and 95.33% respectively as shown in Table 2. A very promising accuracy obtain we MV classifiers deployed on ROIs size (50 x 50) as shown in Figure 4 and observe that random forest classifier performs outstanding among all of these a confusion matrix shows in Table 3.
[39]   In this study MV classifiers employed on human face digital images for recognizing the emotions. Dataset divides into three classes happy, sad, and angry. The various image processing technique is employed and three MV classifiers (RF, Lg, and J48) are tested using cross-validation (10-fold) approach and obtained very promising accuracy 96.33%, 95.67%, and 95.33% respectively. It has been observed RF classifier perform very well among all of these. The obtained accuracy for happy, sad, and angry were 97.16%, 96.83%, and 95% respectively.

VI. ACKNOWLEDGMENT
Authors would like to thank the referee for careful reading and for their comments, which significantly improved the paper.