Research On 3D Facial Expression Recognition Under Uncontrollable Conditions

With the continuous advancement of artificial intelligence technologies,facial expression recognition as a biological recognition technology has received widespread attention in computer vision and pattern recognition areas.Studying facial expressions can make machines understand human emotions more accurately and respond accordingly,thereby promoting better human-computer interaction.At present,2D facial expression recognition technology has gradually matured,but 2D images are susceptible to uncontrollable factors such as illumination and pose variations,which greatly limits the application of 2D facial expression recognition in actual scenes.The 3D face image is robust to illumination and pose variations because it contains facial spatial geometric information,which overcomes the inherent problems faced by 2D face images and has high practicability.This thesis focuses on studying 3D facial expression recognition under uncontrollable conditions.The main research contents include:(1)This thesis studies a method of 3D face reconstruction based on a single image,which is based on the morphable model theory,uses a cascaded convolutional neural network model to regress facial deformation parameters,and then performs 3D face fitting based on the learned parameters.The experiment uses 2D face images with various expression variations and pose variations for 3D reconstruction,which verifies the effectiveness of the method.(2)This thesis proposes a 3D facial expression recognition across pose variations method.Firstly,the 3D face database is expanded by a 3D facial pose augmentation approach.At the same time,rich pose variations are introduced.Then all 3D face models are constructed into grayscale normalized depth images.Finally,a convolutional neural network model is designed as a feature extractor to automatically learn feature representation from depth images to achieve end-to-end classification and recognition.Experimental results suggest that the method is robust to a certain range of pose transformations.(3)This thesis designs an improved convolutional neural network model.The convolutional layer with a kernel size of 5 × 5 is replaced by two consecutive convolutional layers with a kernel size of 3 × 3.Furthermore,the secondary feature activation increases the nonlinear expression ability of the model.The effectiveness of the method is verified by experiments,the network parameters are simplified without loss of recognition accuracy,and the training speed of the model has been increased by 35%.(4)The convolutional neural network model in this thesis uses the cross-entropy loss function and the center loss function as the supervised signal to update the parameters to achieve the goal of maximizing inter-class differences and minimizing intra-class differences in facial expressions,making the model have stronger discrimination ability.The experimental results show that after adding the central loss function,the overfitting problem of the model is alleviated,the recognition rate has improved by about 4%.