Research On Computational Aesthetic Of Image Classification And Evaluation System

The purpose of the computational image aesthetics research is to endow computer withability to assess the aesthetics value of images as human being does. The results can be usedin many fields, for example semantic-based image retrieval fusing the subjective perception,image aesthetics evaluation, image aesthetics prediction and retouching, art works styleanalysis, man-machine interaction, and also design, photography, advertising,etc.In this paper, basing on the digital images, we try to explore the computational method forimage aesthetic classification and image aesthetic evaluation. Firstly, we divide an image intoa whole region and a main region. And then extract the effective low-level visual features andhigh-level aesthetic features. After that, we use the machine learning method to train thefeature data, as a result building an image aesthetic classification classifier and an imageaesthetic score prediction model. Finally, refer to the classifier and prediction model, themachine can automatically evaluate the high aesthetic value or low aesthetic value of animage and predict the aesthetic score of an image.We focus on computational aesthetics of images, and get the following achievements:1、Basing the research on human visual aesthetics psychology, we build an aesthetic imagedatabase which contains human aesthetic evaluation information.2、In order to achieve the subject region as the main region, we combine water-segmentationmethod and the gradient feature to extract the Subject Region of an image as the mainregion.3、We realize the image aesthetic feature extraction, including low-level visual features,high-level aesthetic features and region features.4、We use the machine learning method to train the feature data, specifically use theAdaBoost method to build the image aesthetic classification classifier, and use the SVRmethod to build the image aesthetic score prediction model.5、We finally build Computational Aesthetic of Image Classification and Evaluation System.In this paper，we do experiment on a large number of images which contains aesthetic information. We get an average classification accuracy of77.4%on image aestheticclassification. At the same time, we achieve a correlation coefficient of0.795and Root MeanSquare Error (RMSE) of0.244on image aesthetic score regression prediction. Theexperiment result shows the effectiveness of our method, and the aesthetic results are highlyrelevant to human’s perception on image aesthetic.