The Improvement And The Application Research Of The Facial Expression Dataset In The Wild

As the researchers focus on facial expression research,facial expression recognition gradually become a research hot spot.The disadvantage of existing most expression datasets are that the expression is lack variety,the number of images is not enough;most expression images are collected in the lab.To adapt to expression research,my lab constructed a facial expression dataset with label in the wild(FELW)that the number of images is more(26848 images)and the classes of expression are more(10)in 2016.The FELW dataset have a no high recognition rate,to improve the facial expression recognition rate,preprocess the original dataset to a new database FELW2.0.Then,I have an application research of the FELW2.0dataset using traditional method and deep learning method.Improve the database.First,revise the five parts to eyebrows,eyes,mouth and facial image orientation;next,add the coefficient under the expression label,the coefficient means the judge of the label;finally,extend the people who label to three.The three sets of label will be fused to one set of label by Kappa consistency check,the label and facial image constitute to FELW 2.0.In order to verify the Kappa consistency check whether to work,the three sets of label will be fused to one set,the label and facial image constitute to FELW2.0-1.Then,experiment by the traditional method and the deep learning method on FELW,FELW2.0 and FELW2.0-1.Respectively use hard classification,soft label,soft label and facial parts label experiment.First of all,the traditional facial expression recognition method is adapted to FELW,FELW2.0 and FELW2.0-1 datasets,comparing with experimental results.The result is that FELW2.0 dataset using soft label and facial parts label have the highest recognition rate,the recognition rate can reach 69.09%;to improve the recognition rate,choose front face to experiment,results show that the FELW2.0dataset with front face have higher recognition rate,recognition rate can reach74.91%.Then,the deep learning method is adapted to the three datasets.Three databases respectively use same model experiment,the results show that FELW2.0dataset have the highest recognition rate;Three model experiment are adopted to the same dataset,the results show that model 2 has the highest recognition rate;Usingthe same model on the same dataset,results show that the FELW2.0 dataset what choose front face use two labels have the highest recognition rate,recognition rate can reach 78.22%.FELW2.0 dataset have no normalized operation(image registration),and the image is not collected in the lab but in the wild,so the quality of the facial images are far less to existing most expression datasets.Even so,the recognition rate in the traditional method and the deep learning method can reach more than 70%.