Deep learning based efficient emotion recognition technique for facial images

Facial Emotion/Expression Recognition (FER) is the key technology which is responsible for analyzing facial emotions from digital images in order to reveal information on the emotional state of a person. It is one of the trending research areas in human computer interaction (HCI). This paper is supposed to provide a recent outlook in this arena by combining saliency map and deep learning approaches. As the usage of deep learning-based approaches in computer vision has revolutionized the way such problems are addressed, deep networks like convolutional neural networks (ConvNet) have become the standard design for image recognition problems. The learning through ConvNet needs a large number of images. Apart from the features of interest, most of the time these image data also carry a bunch of nonessential things, which add to the noise in our dataset. Saliency Maps accomplish this task of fixing critical pixels while ignoring the remaining image background. The present paper divides the whole process of FER into three steps: in the first step, image saliency is detected using a saliency map to focus on important regions of interest from the input dataset. In the second step, data augmentation is applied to balance all the emotion images datasets used in testing. In the third step, a deep convolutional neural network is trained using a modified adaptive moment estimation optimizer (M-Adam) to recognize the facial emotions. After this, the proposed technique is tested on the Japanese female facial expression (JAFFE), the extended Cohn-Kanade (CK+), and the face expression recognition plus dataset (FER+) benchmark datasets. The maximum accuracies achieved with the proposed technique are 97%, 99.8%, and 82.7% for the JAFFE, CK+, and FER + datasets, respectively.