Face Super Resolution Reconstruction Based On Conditional Information

The research on face super-resolution using Generative Adversarial Network(GAN)has developed rapidly in recent years and is of great significance in the fields of image reconstruction and security monitoring.The recent concern about racial discrimination in the field of face super-resolution has prompted the problem of face super-resolution: the racial category of the reconstructed high-resolution face image,the facial features are highly random,and even “biased”.This paper investigates and counts the feature distribution of most face data sets in the field of face image super-resolution,analyzes the uneven distribution of ethnicity in the sample,and concludes that there are roughly two reasons for the problem: First,the current most face data sets are more or less imbalanced in the amount of ethnic data,which cause the model to remember this pattern and tend to be more biased towards a specific race when generating high-resolution face images;Second,the race and other facial information of low-resolution face images is often lost,and conditional information(such as race,identity,etc.)is not added to the super-resolution model of the face resulting in uncontrollable race and identity in the image generated by the model.In response to the above problems,this paper mainly made the following contributions:(1)A face data set with a scale of tens of thousands was collected.This data set contained three types of face images of yellow,white and black people,and the category distribution was balanced.This paper verified the advantages of this data set in the data expansion experiment in Chapter 3: The super-resolution model trained on this data set generated more balanced sample results on various races,and solved the problem of "ethnic discrimination" caused by data.The data source and collection and processing methods was introduced in detail in the data introduction section of Chapter 3.(2)A face super-resolution model based on racial information was proposed.By introducing racial condition information,the racial category of the reconstructed face image is controllable.Therefore,high-resolution face images of three races can be generated by inputting any low-resolution face image.In comparison with the classical super-resolution model,the proposed model achieved better experimental results in both objective indicators and subjective feelings,which fully verified the advantages of the proposed model under the control of race conditions.Through ablation experiments,this paper further verified that the Dilated Convolution module and Patch GAN(GAN Based on Patch)module added to the model were the key to the success of the model architecture in this paper.(3)A face super-resolution model based on identity information was proposed.By introducing identity condition information,the reconstructed face image had specific identity information(referring to appearance characteristics).In this paper,the face super-resolution model based on ethnic information was improved,and the face recognition model was used to extract facial feature information to realize the control of identity information.In this paper,part of the experimental results were shown first.Then,by comparing with other super-resolution models,it was found that the proposed model can better recover the appearance features of the original face image under the guidance of identity information.In summary,this paper effectively solved the problem of "ethnic discrimination" through the collected ethnic data sets,and proposed a face super-resolution model based on conditional information,which realized the control of ethnicity and identity information.