Research On Visual Enhancement Of Picture Aesthetics Based On Multi-feature Joint Learning

The aesthetic research of computer vision can be regarded as a reliable materialization process of human aesthetic consciousness.Using the rich information of color appearance and geometry structure of image,model and many other visual media,and with the help of the ability of computer to extract the visual abstract features of data to a certain extent,the corresponding aesthetic tasks are modeled and studied.Among them,face beautification as the most popular research problem has been widely concerned by researchers.In addition,all kinds of image beautification research is also developing.These studies have been widely implemented in many industrial applications.The first task of this paper is to adjust the geometric beautification of human face,that is to improve the beauty of human face by adjusting the proportion of facial features.Most of the existing face geometry beautification and adjustment mostly adopt the traditional machine learning method.This method proposes a face geometry beautification adjustment method based on deep learning to solve the task of face geometry beautification.In fact,in order to get a satisfactory portrait picture,in addition to the post editing of the image,people will choose the best shooting angle in the early shooting.Not only the viewpoint orientation should be considered in face image,but also in other image shooting.For building image,different viewpoints also directly affect the effect of building image.In another work,we propose a method of building image viewpoint evaluation and recommendation,considering the source of image acquisition,to explore the impact of different viewpoints on the image visual beauty.Specifically,this paper mainly includes the following two aspects:1.Geometric beautification adjustment method for human face.This method proposes a GA-GANs network based on deep learning,which can be used to adjust thetask of face geometric beautification.This method is based on the framework of generating confrontation network to model the task of face geometric beautification and adjustment.It aims to let the network learn the mapping method from common face geometric features to high face geometric features,so as to improve the beauty of face.This method adopts the idea of multi-feature joint learning training,and the model contains two generated network branches to learn the complementary information between multi-view features.In feature selection,we select two kinds of features: face key points and geometric distance vector after triangulation,and send them to corresponding network branches for collaborative learning.Finally,the two branches of GA-GANs can independently adjust the face geometric beautification.In this paper,based on the open face data set SCUT-FBP5500,some face images are collected from the network.After a large number of users mark and remove ambiguity samples,it can be concluded that certain aesthetic stability is guaranteed.Compared with the existing face geometry adjustment method based on machine learning,the experiment shows the effectiveness and visual significance of this method.2.The view point aesthetics recommendation method for building image.This method proposes a method of aesthetic recommendation of building image viewpoint.We use the idea of multi-view joint learning in the previous work.The common image aesthetic rules and professional experience skills are comprehensively considered,and the two types of features are combined to learn by using the homology and heterogeneity of the two-dimensional image features and three-dimensional geometric features under the same view point,so as to assist in obtaining the view point aesthetic preference of the building image.Finally,the trained multi-view learner can effectively recommend the view point of the image.This model can also be used when there are only twodimensional image features of buildings or three-dimensional geometric features of buildings,which is in line with the situation that it is difficult to obtain two kinds of features together in reality.The robustness of the model is enhanced.In the selection of multi-view learning methods,we choose CCA and KCCA to compare.Compared with the experiments using only single image feature,geometric feature or feature set without multi-view learning,the comparison results verify the validity and applicability of the two model algorithms.Among them,the KCCA based method is the best in all kinds of feature inputs,and the evaluation results are only inferior to human realjudgment.