Facial Expression Recognition Based On Deep Neural Network

Data is an important section of our lives today,from roads to hospitals,to restaurants,and deep learning is one of the most effective and effiient methods for data pattern extraction.Image detection is increasingly and widely applied for different purposes,such as expert evidence,plant identification,tumor recognition,and facial recognition,among others.In this paper,we focus on the facial emotion recognition,in which the deep learning method still encounters stability and infinite feasibility problems for faces of different races,and the main research contents and innovations arc as follows:Firstly,we introduced the research background the contribution to face detection field.The facial emotion recognition belongs in the supervised learning technique.We proposed a new Bottleneck Feature Extraction(BFE)based Deep Neural Network(DNN)emotion recognition system model to detect the emotion of given images,and the main points was data pattern extraction through the optimal feature selection.In the BFE pre-processing section,we used the Haar classifier for extracting faces from the images,and also used this method to remove the background of the images.After that,we cropped the extracting images by the Haar classifier into a training directory.The aim for Haar was to carry out facial detection with Adaboost and cascading.Secondly,in the Feature extraction of the image,we emphasized on the different variab les,such as color features,texture features,geometrical and shape-based features and topological features.Then we used data argumentation to explain the feature extraction especially through the VGG16 Transfer Learning model for feature extraction.DNN model with five dense layers was used for training the features extracted by VGG16,and the famous Canade-Kohn dataset was used for model training.To evaluate the performance of the model,K-Nearest neighbor and Logistic regression model also was replicated for comparation.Finally,after training on the same dataset,we compared the proposed model with the K-nearest neighbor and logistic regression models.The experimental results showed that our model was more stable and could achieve a higher accuracy and F-measure,up to 98.59%,than other methods.In the future,BFE model is expected to offer robust bottleneck features for complex input images.We also believe that this BFE technique can be used in any circumstance that works with or will work with other images detection field.