Facial emotion recognition (FER) has gained interest and focus over the years. It can be useful in many different applications and could offer significant benefit as part of feedback systems to help train children with Autism Spectrum Disorder (ASD) who struggle to recognize facial expressions and emotions. This paper explores the effectiveness and significance of image pre-processing in Neural Networks on developing suitable models for classification. Transfer Learning using the popular “AlexNet” architecture was used in the development of the model with three different approaches for image inputs. Model performance was compared using accuracy of randomly selected validation set after training on a different random training set from the Oulu-CASIA database and visualizations of predicted areas of importance analyzed. Image classes were distributed evenly, and accuracies of up to 99.90% were observed with small variation between approaches but significant difference in regions of impact. The visualization process highlighted the importance of image pre-processing prior to network training to improve accuracy and eventual efficacy for this application in ASD.
展开▼
