A study on identifying emotional status of rat detecting facial expression in real time with deep neural network

Abstract

The Facial Action Coding System (FACS) is used to better understand the action unit, which is the movement of the muscles, to identify the facial expressions of humans. The emergence of an active appearance model based on masking in human faces helped to facilitate recognition of FACS. However, it is difficult to clearly figure out how the expressions of animals would indicate their specific feelings. Therefore, we look back on the methods applied to estimate the feelings of humans from human facial expression and see whether similar methods could be used in rats or not. First, we need to know exactly how the processes of identifying human action units or emotions. In order to do this, we propose applying the traditional methods of detection, active appearance model (AAM) tracking, and the use of Deep Neural Network. First, features were extracted using the AAM method based on the CK+ database. The model was created by reducing the features extracted here and learning from the SVM. Compared to the database results presented by CK+, the feature reduction was mostly more accurate in the AU segment, while the feature reduction in the Emotion portion was lower. It also built a neural network model with a DenseNet structure and extracted and learned data from EmotioNet, a large database. An average of about 80% accuracy was shown for action units. Finally, we investigated emotion estimation studies of rats. A study was conducted to find the pain of mice using the Rat Grimace Scale, a pre-determined unit action after suffering in a typical physical/chemical manner, and a method was presented to focus on the color of the ears from sparingly to find a positive expression. These are action units that are visually distinct and have protocols presented. The fine-grained unit of behavior toward humans has shown some discernment, so it is expected that the results will be valid if only a database targeting mice are available.