Adaptive Facial Expression Recognition With Its System Design And Implementation

Facial expression recognition often referred to as FER is an important branch of face recognition,which is also a cross research field involving many disciplines including psychology,biology and computer science.Although there are many FER algorithms based on the deep neural network,the recognition accuracy and practical application effect of them are not ideal.There are three main problems: 1)The quality of existing public FER datasets is not high,either the annotation is detailed but the number of effective data is small,or the dataset is huge but the image quality and labeled information are not guaranteed.2)The standards of various FER datasets are uneven,and the improvement of one algorithm is often only effective on a certain dataset,which is not conducive to the parallel research and comparison of researchers.3)Many FER algorithms are still under the research stage.They do not pay attention to the difference of face images between the real scene and the natural scene.Thus,the actual application effect may be greatly reduced.The purpose of this paper is to sort out the existing FER datasets,increase the total number of images in dataset as much as possible,and get a new large-scale hybrid facial expression recognition dataset(LHFER)by using data cleansing technology,on the basis of ensuring the quality of face image and the accuracy of labeled information.Meanwhile,considering the imbalance of the number of expression categories and the large difference of face image resolution,we select and create several sub-datasets to better adapt to various complex situations in the real application scenarios.For our new dataset LHFER,this paper proposes an improved covariance pooling method for facial expression recognition.Various baselines and advanced FER algorithms are also reproduced and retrained on LHFER to explain the effectiveness of the proposed algorithm.In addition,the algorithm trained on LHFER will verify its generalization and robustness in real classroom scenarios.Finally,the system design and application implementation will show the applicability of the adaptive FER algorithm based on LHFER.