Estimation of Emotion using CNN

Abstract

The estimation of human emotion is still a challenging task in computer vision. Breakthrough in this direction can aid in humanoid robots, can make blind people understand the human behaviours. In order to predict human emotion, detecting the human faces in the image is the primary task. Viola jones and Adaboost algorithms are used for this purpose. The face is detected using the frontal face haar cascade classifiers. The disadvantage with the frontal face is that the face is detected only when the complete front face is there in the image But it is unavoidable because the emotion can be predicted only if the whole front face was observed. A convolution neural network is implemented using Keras and tensor flow in order to detect the emotion. The CNN is trained using 28709 images and tested using 7178 images of Kaggle dataset. The trained model successfully detect the emotions of happy, sad, Disgust, surprise, angry, and neutral. The testing dataset is again splitted into 3589 for validation and 3589 are used for testing. The validation dataset is used in predicting the training accuracy vs validation accuracy and Accuracy, Precison, Recall and Fscore are calculated to evaluate the model using the testing images.