Facial Expression Recognition in Static Images for Autism Children using CNN Approaches

Humans express their emotions through facial expressions which is an effective tool in communication. Facial expression recognition (FER) such as non-verbal communication is one of the vital activities when interacting with others. FER is the topic of a current study in the field of artificial intelligence, with several recent experiments utilizing convolutional neural networks (CNNs). This paper proposed a system that classification FER based on static images using CNNs. It used a histogram of the oriented gradient (HOG) for information extraction, then applied a support vector machine to detect eyes and mouth; finally, FER depended on CNN algorithms to achieve greater accuracy and significantly shorter training times. To train the model, where the dataset was collected from standard databases, the expanding KDEF dataset contained 1,960 images. Tested our model on expanded KDEF and another type of KDEF which contained 2,934 images. The system was examined and evaluated CNN model results showed significant success in the accuracy of FER was 97.48% during training phases, and a validation accuracy was 92.66% accuracy of the CNN model testing. The accuracy of testing for the real dataset from Alnoor Specialized Center containing 1000 images for classifying facial images of autistic and non-autism children using CNN is 99.78%.