Research On Ultra Low Resolution Face Reconstruction Based On Multi-Output Regression

Super-resolution face reconstruction technology is to reconstruct high resolution imagefrom low resolution face image, it has important practical significance in public securitysystem, monitoring system and electronic commerce system. Ultra low resolution face imagewhose scale is less than24（24×24） pixels is serious lack of face image information, it’sreconstruction has more values not only to machine automatic identification but also toartificial discrimination, but only a few papers research this problem. This paper studies thereconstruction for positive face image whose scale is less than24（24×24） pixels to faceimage whose scale is24（24×24） pixels as the core content.According to the reconstruction in the existing systems whose model is big andreconstruct speed is slow as the actual situation, this paper proposes a piece-division multipleoutput regression algorithm based on the bayesian multiple adaptive regression splines（MARS） model to make ultra low resolution face data’ regression. This paper studies thelow resolution face image reconstruction, which has less data, using the regression method toreconstruct is the most appropriate way with small model, fast reconstruction speed andaccurate results as its advantages. The bayesian MARS method uses MARS as regressiondiscrimination function, and decides MARS parameters through the bayesian learningmethods and Markov Chain Monte Carlo method. Compare with the traditional certaintylearning method, the random study method enhances the robustness of the parameters in thebasis of keeping the accuracy of acquiring parameters. Accurate MARS parameters ensure ahigh accuracy of reconstruction of this method. The piece-division regression method whichdivides both low resolution face images and their corresponding high resolution face imagesinto some same size pieces can increase the number of models, reduce the size of each modeland accelerate the reconstruction speed. The experiments prove the reconstruction accuracyfrom error and identification accuracy two aspects in this paper.Meanwhile, the low resolution face detection and alignment methods are also researchedin this paper, the method based on Haar feature and AdaBoost classifier is choosed to detectface image for its effective to low resolution images. And a alignment method based on agradient image is proposed in this paper, it makes low resolution face alignment moreaccurate. After aligning the images, the reconstruction accuracy can get increased. The experiments prove the reconstruction accuracy after alignment has increased significantlycompared with the reconstruction accuracy without alignment from error and identificationaccuracy two aspects. So this paper realizes the detection, alignment, reconstruction of a realimage in the face practical system and practical application of face reconstruction.