Image Aesthetic Quality Evaluation Based On Convolution Neural Network

Along with the development of Internet and mobile device such as mobile phones,the number of image increases rapidly and every day a lot of images are loaded on line.To help people exhibit the high aesthetic quality images and explore the perception of aesthetic,people pay more and more attention to image aesthetic quality evaluation.Aiming at the problem of small-scale data and lacking researches on the factors of affecting image aesthetic quality,this paper mainly uses convolution neural network(CNN)to evaluate image aesthetic qualities.First,we propose embedded fine-tune based on image content.Second,we build a dataset for Image Aesthetic Factors Analysis(IAFA)by using a novel way of using computers to help people score images.At last,we use convolution neural network to study on the factors of affecting image aesthetic qualities based on IAFA.Embedded fine-tune is a way of using twice fine-tune continuously.The whole data is used in the first fine-tune and part of the data is used in the second fine-tune.The experiment results on Photo Quality show that the categorization accuracy of embedded fine-tune is higher than the existing approaches.Solve the small-scale data problem well.At last,we use the trained models to evaluate image aesthetic qualities.Divide 26 kinds of features extracted by our previous work into composition,depth of focus,brightness and color.Score the four image attributes by compromising the weak label of computers and strong label of human and build IAFA with 24631 images.We hope that IAFA can encourage further research on image aesthetic quality evaluation.We also perform extensive benchmarking analyses on this new data set using the single task learning and multiple tasks learning of CNN.