Research And Optimization On Loss Function Problem Of Deep Face Model Based On Data Cleaning

As the size of face data sets continues to grow,researchers have designed a variety of faster and stronger face recognition networks for face recognition,but we still know less about the source and consequences of the noise in existing data sets.Since the model is based on data fitting,the quality of the data will directly a?ect the final result of the model.For small data sets,we can clean the noise by manual labeling,but in the face of large-scale data sets,these problems can not be completely solved by manual.And the cleaning data will also reduce the scale of the data set to a certain extent.To this end,this paper has carried out in-depth research around data cleaning and the use of improved loss function to enhance network learning information.A streaming data cleaning method based on hierarchical clustering is proposed to solve the problems encountered in actual scenarios.First,we analyze the noise problems of di?erent kinds of data sets in detail.After that,from the structural aspect,we designed the cleaning method into a streaming structure,and the data passed through each cleaning module in turn,and finally a clean data set was obtained.From the content aspect,we use the idea of hierarchical clustering to clean the data set with high precision through three-layer clustering.A loss function method based on random noise is proposed to solve the problem of data set size reduction caused by data cleaning.First,we analyzed the possible forms of noise.After that,we added random and small noise to the facial features of the network output,and used the LFW and YTF test sets to evaluate the trained models.Finally,the accuracy of the model under di?erent noise combinations is obtained through experiments.The experiment proves that the recall method of the cleaning method proposed in this paper reaches 99%,which can solve the problem of data noise.The model based on the loss function of random Gaussian noise is tested on LFW and YTF,and the accuracy is improved by 0.15% and 0.27% respectively.