Facial Expression Recognition Based On Double Weber Local Descriptor And Deep Belief Network

As an important way of expression of human emotions,facial expression is a important way by which human emotions can be identified and understood by machine in the human-computer interaction.Especially,with the wide application of human-computer interaction in the life,higher demand for machine's automatic facial expression recognition technology was presented.Facial expression feature extraction is a key step in the process of expression recognition.And the results will affect the recognition accuracy and algorithm performance directly.However,there are still problems of the current algorithm of facial expression feature extraction such as,(1)the feature of the image details is not comprehensive and accurate;(2)There are interference caused by changes of light and noise,and other independent information related to identification;(3)Facial expression extraction time is a bit long,and so on.This paper focuses on the expression feature extraction and improves it.First,In order to optimize the texture information extraction of traditional Weber descriptors in detail,The utilization of adjacent pixels of the center pixel was optimized for improving the Gradient Orientation.Secondly,in view of the fact that local texture information is insufficiently extracted and computationally intensive,when Deep Beef Network(DBN)is used for extracting the feature of images.In this paper,Double Weber Local Descriptor(DWLD)is used to extract the image characteristic data as the initial feature,and as input for Deep Belief Network for the higher level of feature abstraction.The fusion characteristics has a more comprehensive image characterization capabilities,and the recognition accuracy and algorithm efficiency are improved.The main work of this paper is as follows:(1)The research results of facial expression recognition in recent years at home and abroad had been summarized in this paper,and the algorithms of expression recognition were expatiated.The algorithms of feature extraction algorithm used commonly and the corresponding improvement direction are introduced.(2)In view of the problems such as inaccurate characterization,lack of validity,when feature of the expression image detail was extracted by the existing expression feature extraction algorithm,Double Weber Local Descriptor is proposed in this paper to strengthen the characterization of the local details of the image.In the calculation of the Gradient Orientation,the DWLD increases the utilization rate of the adjacent pixels of the central pixel,which makes spatial distribution is optimized and texture information more discriminative and comprehensive.(3)In order to improve the comprehensiveness of the expression image features,a new expression recognition algorithm of fusing Double Weber features and Deep Belief Network was put forward in this paper.The problem of the lack of representation in the overall structure of the image when Double Weber feature was used for feature extraction is improved,through the powerful abstract ability of Deep Learning to extract the overall information of image.Firstly,the initial feature is extracted by using Double Weber Local Descriptor.Secondly,the secondary feature extraction is carried out by initial feature as input into the Deep Belief Network model.A higher level of abstract features is extracted by the result of the fusion,which can combine the advantages of local features and global features.Because the more representative characteristics of the images was used to be learned and trained by DBN,the redundant information such as light,noise which have nothing to do with expression recognition is reduced.So the identify speed has been significantly improved,and classification effect is better with the BP classifier.(4)The standard facial expression databases are used in this paper for independent testing,and the experimental results showed that the proposed method improved the facial expression recognition rate effectively.Compared with the use of the Deep Belief Network alone.the total time consumed s reduced by 13.89% by the fusion of the DWLD and DBN.Therefore,the computational complexity of DBN is reduced,which improved the recognition efficiency significantly.