Multi-view Facial Expression Recognition Based On Dense Capsule Network

At present,most facial expression recognition methods are performed on frontal faces,but frontal facial expressions are often an idealized situation.In real life,the capture device may capture expression data from multiple angles.The existing multi-view facial expression recognition algorithms are mostly based on convolutional neural networks.Convolutional neural networks cannot express the spatial hierarchical relationship between object parts.Therefore,these models generalize poorly across novel views that are not included in the training set.In view of the fact that the capsule network can model spatial hierarchy,we propose a more robust DenCaps model based on the capsule network.The DenCaps model incorporated the Dense Block module of DenseNet network and capsule network.The Dense Block learns more features through feature reuse and bypass settings.The capsule unit in the model uses vector instead of scalar storage entity features.Compared with the traditional convolutional neural network,it can encode the pose,position and other parameters of the object,and achieved good performance in multi-view object recognition.We tested the DenCaps model on the FERA2017 dataset and compared it to other recent network models.The experimental results show that the DenCaps model achieved an F1 value of 53.9 on FERA2017,which is higher than other network models.In addition,we performed a cross-pose experiment based on DenCaps.The results show that DenCaps can detect facial action units from facial images from novel perspectives.Finally,we perturbed the single dimension of the capsule in the DenCaps model and reconstructed the image to visualize what the capsule learns.Visualization indicates that the capsule is capable of storing instantiated parameters such as pose,head size,AU intensity and illumination.