Ensemble Deep Neural Networks For Facial Expression Recognition

Deep learning has dramatically improved the accuracy of face recognition in recent years.To achieve higher recognition accuracy,ensemble learning can be applied to these varied deep learning algorithms.Facial expressions convey useful information that is difficult for traditional classifier to detect.There are many problems in facial expressions as like low resolutions,occlusions,different lighting and positions etc.Often,the human can't recognize them due to the expressions are categorized very poorly.In addition,the categorization of a facial expression is not always universal.For instance,a smile does not always mean happiness.Facial expressions often depend on cultural constructions.However,the ability to recognize them can lead to more responsive and intelligent systems that might improve the user experience.In order to enhance classifiers performance and reduce the error rate in facial expression recognition,many works have been developed such as deep learning.Sometimes deep learning can't recognize facial expressions for several reasons like implementation can be a complex and difficult task,and high-quality datasets can be hard to find.Moreover,the performance of deep networks depends heavily on the large number of labeled samples.In this study,novel methods based on convolutional neural network and ensemble deep networks are presented when only very small number of samples are available.These methods are a Multi-View Convolutional Neural Network(MVCNN)and an Ensemble Transfer Learning Network(ETLN).First,facial images are downsampled to different scales and upsampled as multi-view samples.Then,a MVCNN is constructed with twin structure.After one channel is trained by single view samples,the parameters are transferred to another channel for fine tuning using different samples.Furthermore,we enhance the results of the MVCNN method using ETLN technique based on transfer features and ensembled deep models.First,DCNN network,which is designed from scratch,is used to recognize the facial images.Then,VGG16 and VGG-face networks are employed to build other proposed individual classifiers Transferred VGG16 and Transferred VGG-face.The individual models are aggregated to build the Ensemble Transfer Learning Network(ETLN)that is applied in different cases.Moreover,ETLN improved the performance clearly after the weight analysis is applied where the optimal combination of features is the one that achieves the highest accuracy.These methods have been implemented and tested on FER2013 and RAF-BASIC datasets,and the recognition results are evaluated by a comparison with those of its counterparts.The experiments results show that the proposed methods can provide more accuracy where MVCNN classifier achieves an accuracy of 72.27%on the test set of FER2013,and ETLN classifier achieves an accuracy of 74.12%,85.94%on FER2013 and RAF-BASIC,respectively.