Emotion-Detection

Facial Emotion Recognition

• Abstract
• Methodology
• [Results]
• References
• Future Scope

Abstract

Facial expression for emotion detection has always been an easy task for humans butachieving the same task with a computeralgorithm is quite challenging. With the recent advancement in computer vision and machine learning, it is possibleto detect emotions from images. In this paper, theaim of this project is to develop a Facial Emotion Recognition model using ConvolutionalNeural Network algorithm that can help for real time emotion detection. For the specific use cases like pain detection in healthcare, or individual emotion monitoring systems while driving the features that constitute for the emotion can also help to improve the accuracy of the recognition. In the ERmodel, expressional vector (EV) is used to findthe fivedifferenttypes of regular facial expressions.The two-level CNN works in series, and the last layer of perceptron adjusts the weights and exponent values with each iteration. ERdiffersfrom generally followed strategies with single-level CNN, hence improving the accuracy. Furthermore, a novel background removal procedureapplied, before thegeneration of EV, avoids dealing with multiple problems that may occur (for example distance from the camera). ERwas extensively tested with images using extended Cohn–Kanade expression, CMU datasets. The ER model with more training can be used for emotion detection to be useful in many applications such as to alarmindividuals while driving, pain detection in the hospitals. Addition of features behind the emotion like Heart rate, ER, Neuro monitoring can help to draw a better conclusion in Emotion recognition.

Methodology

Starting the project with the Emotion Recognition System and will update the models regarding to pain specific data later on. The proposed method has been shownin Fig.1. It’s an amalgam of various techniques for extraction of features from image frames for automatic detection of pain. The details of all the steps are showed in the architecture. Facial Region Extraction, Iterative Shape Alignment Using Procrustes Analysis: The shapes of the faces are iteratively aligned so as to remove all geometric disparity i.e. scale, rotation andtranslation of the vertex location with respect to the base shape [8][9]. The alignment of multiple shapes is based onaligning pairs of shapes where one of them is a reference frame. For this Generalized Procrustes Analysis (GPA) isperformed which consists of sequentially aligning pair of shapes with using the reference shape (the mean shape)and align the others to it. At the beginning of the procedure, any shape can be selected as the initial mean. After thealignment a new mean is recomputed, and again the shapes are aligned to this mean. This procedure is performedrepeatedly until the mean shape doesn't change significantly within iterations.Texture WarpingAfter aligning all the shapes, warping of texture is carried out using a piece-wise affine warping to normalize all non-rigid texture samples with respect to the base shape. A piece-wise affine warping is a texturemapping procedure where convex hull of the mean shape and the sample shapes are partitioned using Delaunaytriangulation [3][4][10]. Then each pixel from the training image, belonging to a specific triangle, is mapped into the respective destination triangle in the mean shape frame using barycentric coordinates with bilinear interpolation correction.

3.2. Feature Extraction using Gabor FilteringAfter extraction of facial regions, the next step is facial feature extraction to produce feature vectors. Gabor filters(also called as Gabor kernels) is an significantmethod in the areaof image analysis because of its optimumlocalization properties in spatial and frequency domain. It is used to extract the changes in facial appearance asa set of multiscale and multi-orientation coefficients. In the spatial domain, a 2D Gabor filter is a complexexponential modulated by a Gaussian function.

Architecture overview

![Gabor Filtering]

Model specifications :

• OPEN CV Face recognition
• embedding size = 200
• batch size = 2096
• numner of epochs = 15
• time taken for each epoch = 1300 sec approx
• CNN Model
• Data set :CNN Face data set

References

Future Scope

1. Training for more epochs
2. Increasing batch size
3. Implementing Attention Mechanism
4. Extending it to Pain Specific data