Research On Image Aesthetic Evaluation Using Image Scene And Saliency Information

With the continuous progress of information technology,people come into contact with a large number of images every day.The form and content of the image become more and more diverse.The purpose of the computational image aesthetics research is to use computer simulation of human visual and aesthetic thinking to make judgments on the degree of image aesthetics.Therefore,the study of image aesthetic evaluation gradually aroused researchers' interest.The results can be applied to image quality assessment,photography aesthetic correction,art painting analysis and other fields.At present,there are two main aspects of the research on image aesthetic evaluation.One is how to obtain a comprehensive expression of the aesthetic value and strong robustness image feature,and the second is how to build and design aesthetic evaluation model.This paper mainly focus on the problem of automated evaluation of image aesthetic.Refer to the classifier and prediction model,the system can classify images into high aesthetic value or low aesthetic value categories and predict the aesthetic score of an image.We mainly achieved the following innovations:1.This paper proposes a method of image aesthetic evaluation that fuses image scene information.According to the different types of image scene with different aesthetic values,the new method comprehensively considers the difference of different types of images,and use a novel convolutional neural network structure(Multi-scene Deep Learning Model,MSDLM)to learning different types of aesthetic features.In addition,in order to reduce the impact of the small amount of training data and noise sample,preprocessing and pre-training are taken into account.From the experimental results,the improved deep convolutional neural network obtains better classification results.2.A comprehensive image aesthetic evaluation model is designed,which includes two tasks: image aesthetic classification and aesthetic score prediction.Considering the aesthetic value of the image salient region,this paper designs a parallel deep convolutional neural network.One of them is used to learn the global aesthetic features,and the other is used to learn the local aesthetic features.Then we combine these networks to get an output.In this paper,aimed at the existing problem of image aesthetic evaluation research,we propose new idea and model design.The new method has achieved better results than other methods,the output of the system is consistent with the results of human aesthetic perception.