Research On The Evaluation Of Personalized Image Aesthetics Quality

The purpose of personalized image aesthetics evaluation is to predict the individual's unique aesthetics of images,and it has good application prospects in image recommendation,image editing aesthetics and other industries.In recent years,research in this field has gradually emerged,but it still faces the following challenges.First of all,from the perspective of personalized aesthetic data,due to the complexity of the aesthetic evaluation problem itself,the manual labeling of data is very time-consuming and laborious,resulting in the small scale of currently available personalized aesthetic evaluation data sets(especially scoring data sets)and evaluation The data often has certain deviations.Secondly,from the perspective of the scalability of personalized aesthetic quality evaluation methods,some methods lack a mechanism to collect the aesthetic preferences of new users,and can only be evaluated for users who have been modeled;while other methods can use reinforcement learning and other methods.Establish aesthetic preferences for new users,but they still have their own limitations.Finally,most methods have a single evaluation index and only evaluate from the characteristics of the image itself and the overall layout,without considering the subject's important influence on aesthetic evaluation.In response to these problems,this article studies the existing personalized image aesthetics quality evaluation methods,the main research contents are as follows:(1)A personalized image aesthetic quality evaluation method based on the attention mechanism is designed,and the attention mechanism is introduced to the original personalized image aesthetic quality evaluation method.Add the subject's significance factor to the scoring process.The addition of the attention mechanism greatly improves the accuracy of the model.Experiments on the Flickr data set show that compared with other traditional methods,this method improves the SROCC index by 3%.(2)A new user aesthetic preference expression method based on preferred image pairs is designed,and an interactive user aesthetic preference classification mechanism based on a decision tree is proposed based on this method.The user aesthetic preference based on the preferred image pair does not require the user to spend time on accurate aesthetic score calibration.It is a user-friendly,efficient,and explainable way of expressing user aesthetic preferences with clear and uniform rules.This paper builds a user classification decision tree through training based on this preference expression.Only a small amount of simple feedback is required to model new user preferences.The above strategies effectively solve the aforementioned problems of personalized data,overcome the difficulty of modeling new users to a certain extent,and improve the scalability of the method.Experiments on the FLICKR data set show that the method is accurate and effective in user personalized preference classification.(3)A multi-task-based personalized aesthetic scoring system,MTPAA,is designed,,which integrates multiple aesthetic evaluation labels such as general aesthetic evaluation grades,image styles,universal aesthetic evaluation distributions,universal aesthetic evaluation scores and personalized aesthetic evaluation scores as output tasks In a complete system,through a staged multi-task training method,an efficient and comprehensive aesthetic quality evaluation is realized.In addition,the aforementioned interactive user aesthetic preference classification mechanism based on decision trees is incorporated into the system,which improves the scalability of the system.Experiments on data sets such as Flickr and AVA show that the SROCC indicator shows that the system not only has a slight improvement in performance compared with other methods on each task,but also achieves the diversity of aesthetic evaluation results and expands the network's performance.The scope of application scenarios,and the adoption of a staged multi-task implementation method greatly reduces training costs.