| In recent years,with the continuous progress of science and technology,the field of artificial intelligence has also made rapid development.With the widespread application of artificial intelligence technology in various aspects such as human life and learning,life and health,and travel safety,people have put forward higher requirements for its application ability.The research on artificial intelligence in the field of facial expression recognition is receiving increasing attention from scholars due to its unique application value.Because facial expression recognition is an interdisciplinary subject that spans computer,artificial intelligence,biological neurology and other complex disciplines,it has great application value in machine vision,life safety,human-computer interaction,safe travel and many other fields.Facial expressions are the most direct information transmission carrier for human emotional information,which can express human emotions in real-time.However,the prerequisite for facial expression recognition is to first accurately capture and locate the face,and then preprocess the collected images,in order to better complete the extraction of local facial detail features in the later stage.Feature extraction is the focus of this research topic,and the quality of the feature extraction work will directly determine the final result of facial expression recognition.Therefore,this article proposes a method that combines traditional feature extraction methods with deep learning to complete this important task.The specific research methods are as follows:1.In response to the difficulty of extracting deep and multi-dimensional detail features from facial images using traditional methods,this paper proposes to use the LBP histogram method to extract texture features,which serves as a layer of feature channels for multi-channel feature fusion;Use the edge detection Canny operator to extract the edge features of the image,and use the extracted edge feature image as another feature channel;Perform channel fusion between the above two special images and the original image.Finally,the images with three channel features are preprocessed and input into the already built network model to complete model training and recognition classification.To solve the problem of difficulty in extracting image features using traditional methods2.In order to address the common issues of weak robustness and weak generalization ability in mainstream models,this paper proposes a separable convolutional network based facial expression recognition method improved on the Inception Module.Among them,the model simultaneously uses four improved Inception Modules to undertake the main task of feature extraction in the model framework.Each module contains relatively complete structures,such as pooling layers and convolutional layers.At the same time,this model adopts a diversified combination of asymmetric small convolutions to replace the idea of large convolutions,which can maximize the height extraction of local details in the image by verifying small convolutions.Each module can complete feature extraction work to the maximum extent for different image regions,and ultimately complete accuracy testing on the public dataset FER-2013.And through comparison with mainstream models,it has been proven that this model performs better in deep feature mining and weakening the influence of lighting,and the network model has stronger robustness and generalization ability.3.Considering the practical application of network models and the current issue of safe travel,in order to solve the safety problem of drivers experiencing emotional fluctuations and causing serious losses due to dangerous driving behaviors influenced by the external environment,this model has been tested through practical scenario experiments.At the same time,in order to verify the stability of this model,experiments were carried out on the self-made data set to complete the accuracy test,and the final data proved that this model was stable and accurate.The model has been verified to have strong robustness and recognition accuracy through practical testing. |
