| Facial beautification is a novel computational photography technology used to improve the visual effects of portraits.It has been intensively applied in advertising design,short video live streaming and social sharing platforms.In the actual use of beautification technology,in order to achieve the desired visual effect,it is necessary to continuously adjust technical parameters.This process is time-consuming and laborintensive and cannot be applied to live broadcasts of dynamic changes in the environment at any time.However,the aesthetic quality evaluation of facial beautification images has not yet been widely used.Based on this,this article first studies the subjective quality assessment of facial beautification images;Then,based on the research experience of humans in perceptual psychology,arts and other fields and the principles of facial beautification technology,two effective no-reference quality evaluation methods are proposed;Finally,based on the evaluation methods proposed in this study,a face beautification system based on image quality evaluation was designed.The main contents of the thesis are organized as follows:1.In order to make up for the blankness of the face beautification image quality database,a face beautification image(FBI)database is constructed in this paper.First,collect 25 face images without occlusion,heavy makeup,and relatively uniform beauty;Then,six common beautification methods are used,and each method is set to four levels to beautify each image;Then,30 experimenters performed subjective evaluation experiments on a GUI interface built on MATLAB;Finally,the confidence interval is used to remove outliers in the subjective data,and the average value of the remaining data is used as the mean opinion score(MOS)of the image.By analyzing the standard deviation distribution and histogram of the MOS,it can be known that the MOS value of the FBI database has good dispersion and meets the criteria for database construction.2.A no-reference quality evaluation method for face beautification image based on portrait aesthetics cognitive psychology is proposed.According to the conclusions on the study of facial beauty in disciplines such as psychology,art,and sociology,we can see that the attributes of facial features are completely different,and there are also large differences in their aesthetic standards.Facial beautification is based on these conclusions.It is a technique to change the attributes of the face through image processing to improve the visual quality.Based on this,this thesis come up a noreference quality evaluation method for facial beautification images based on cognitive aesthetic psychology of portraits.First,the face image is decomposed to obtain the skin area,eye area and mouth area;Then quantify the aesthetic characteristics of these areas according to portrait psychology,including the three attributes of skin aesthetics: color,light and smoothness,aesthetic differences caused by gray differences between eyes,mouth and surrounding areas,and image clarity;Finally,a support vector regression(SVR)function is used to train a facial beautification image quality prediction model.The test results on the FBI database show that this method has a high consistency between the predicted score and the MOS value,and the performance is better than the seven main stream no-reference evaluation methods.3.This chapter proposes a no-reference evaluation method for face beautification images based on deep learning.Due to the subjective nature of the human visual system for facial beautification image quality,based on portrait aesthetics and art,the method of extracting aesthetic features has achieved some results,but the extracted features are generally low-level.Facial beauty is energetic and insufficient,and it is difficult to break through in prediction accuracy again.CNNs learn low level features into high level abstract features by learning the human way of thinking,which can effectively improve the expression of abstract features of images.Based on this,a no-reference quality evaluation method for face beautification images based on deep learning is proposed.First,the face image is decomposed into a texture layer,a brightness layer,and a color layer;then these layers are decomposed on the SCUT-FBP5500 database and then input to the Cascaded Convolutional Neural Network(CF-CNN)To perform pre-training,and finally Fine-tuning on the FBI database to optimize the network.The experimental tests evidence that this algorithm can more validly assess the quality of facial beautified images.4.A face beautification system based on image quality evaluation was designed.The research of face beautification image quality evaluation methods is designed to guide the selection of beautification technical parameters.In order to visually show the performance and practical value of the proposed method,a face beautification system based on image quality evaluation is proposed.The program software system is compared of a face area recognition module,a face beautification module and a quality evaluation module.After embedding the image quality evaluation method into the image beautification,it realizes the function of guiding the selection of the beautification technical parameters.The performance of the quality evaluation method in this paper is further demonstrated by practical application.The paper has 26 maps,9 tables,and 102 references. |
