Face Orientation Detection Based On Deep Learning And Its Application

Technology applications about human faces have become more and more common used in today's computer vision.In addition,studies on face detection,face recognition and facial beautification become increasingly further and mature.However,technology always encounters problems in practical application.For example,if there is not a little bit of deflection of human faces in the input image,most of the face detection applications will fail to make a correct detection,unless the deflection of human faces has been adjusted within reasonable bounds by manual operation before input.Moreover,when doing researches on facial beauty evaluation,if the features used in the training process could not keep intact during the rotating,the deflection of human faces will have a bad impact on the evaluation.Consequently,it is very necessary to embed a facial orientations detection module in the system of face detection,face recognition and facial beautification.This present paper has raised a method of faces orientations detection based on Deep-Learning,and this detection model has been put into application in “How-beauty” based on iOS,which is a face beauty evaluation system developed by ourselves.Major work of the research in this present paper and its contributions are listed as follows:(1)I utilized Convolutional Neural Networks in the classified-learning of the faces image,which was one of the four orientations {0°,90°,180°,270°}.And the accuracyobtained was 99.12% on our test set by the FOC-CNN model.(2)For solving the problem of incorrect evaluation caused by the deflection of humanfaces in facial beautification,I utilized Convolutional Neural Networks to realize theregression prediction of face orientations within the range of [-90°,90°].After thetraining a model called FOR-CNN was available and it could be used to predict theface orientation within the range of [-90°,90°].The absolute mean error was 9.25.(3)This present paper also predicted the face orientation within the range of [0° 360°].And this training mission was divided into two sub-missions.One was to classify oneof the four orientations{0°,90°,180°,270°} and the other was to achieve theregression prediction within the range of [-45°,45°].Both of the sub-missions were asthe Deep-Learning target during the training process.With the help of multitaskingConvolutional Neural Networks model,we could then obtain the FOMt-CNN modelafter the repeated iterative training.Finally,the model's absolute mean error was 15.76.Through the comparison of the experimental results,we found that dividing thetraining mission into two sub-missions work more efficiently,and this multitaskingtraining mode could provide a better Convolutional Neural Networks model.(4)We developed a face beauty evaluation application called " How-Beauty " which wasbased on the system of iOS.And this present paper had introduced its design of theprogram and the process of the implementation.Compared with other relative iOSapplications," How-Beauty " obtained higher accuracy of face beauty evaluation andbetter interaction.