Research On Personalized Image Aesthetic Enhancement And Quality Evaluation

In the field of computer vision,there are two main aesthetic research tasks: image aesthetic quality enhancement and image aesthetic quality assessment.Image aesthetic quality enhancement is what gives people the feeling of "beauty" by emphasizing or changing some characteristics of the image purposefully in enhancement model and improving the visual effect of the image.Image aesthetic quality assessment is to evaluate the aesthetic quality of the image by building a high-performance automatic image aesthetic evaluation model.At present,most of the aesthetic models are built based on open data set.But,the aesthetic related research tasks have strong subjectivity and different individuals have different aesthetic perception.These general aesthetic models ignore the consideration of personality and subjectivity.In the research of image aesthetic quality enhancement,this paper obtains the personalized aesthetic enhancement effect of image by improving the aesthetic quality of image,according to the rules of users' aesthetic preferences.In the research of image aesthetic quality assessment,the model in this paper can modify the evaluation result of the algorithm itself by introducing the users' interaction means,formatting the the subjectivity model of user evaluation.And then,the accuracy of aesthetic quality evaluation model is improved.The main contents of this paper are as follows:(1)In this paper,we build a user-friendly data collection platform on web pages,collect 60 users' personalized data and format a personalized data set with 3768 images,which integrates aesthetic evaluation and aesthetic enhancement and provides data basis for personalized research.(2)This paper proposes a personalized image aesthetic enhancement model based on reinforcement learning.Aiming at the personalized problem of image aesthetic enhancement,this paper describes the process of user's retouching image as a sequential decision-making problem,and uses Actor-Critic strategy of reinforcement learning to solve it.Finally,we obtain the personalized enhanced image guided by user's aesthetic preference.The experimental results show that the aesthetic enhancement algorithm proposed in this paper can not only improve the aesthetic quality of the image,but also the aesthetic effect enhanced can meet the user's personalized requirements.(3)In this paper,a framework of aesthetic quality evaluation based on personalized image aesthetic enhancement is proposed.The framework consists of three modules: Firstly,by using the optimized Alexnet network,we extract the aesthetic features of the specified image,which is used to match the similar images from the public database.And then,we obtain the first evaluation model based on the personalization through the personalization ranking algorithm;Secondly,by introducing the rearrangement and editing operations in the user interaction process,an iterative aesthetic evaluation model combined with image enhancement is constructed,which saves the time cost of user interaction,increases the user's subjective initiative and makes the expression of aesthetic subjectivity more accurate.Finally,the user's personalized aesthetic distribution is generated.And we conduct a visualized research on the personalized and subjective of user's aesthetic evaluation by using a variety of aesthetic attributes.