Research On Social Image Aesthetic Classification And Enhancement Algorithm

With the development of computer technology, smart phones and mobile internet,acquisition of images became more and more convenient. In social networks, people prefer toexpress ideas and states with images. The real-time communication on the internet alwaysmake users to neglect the aesthetics of images when sharing them. However, people alwaysshow more interest in images with higher aesthetics, which could bring not only more viewers,but also more benefits for social internet sites. Thus, it’s necessary to rank images accordingtheir aesthetic value and enhance aesthetic value in them. Algorithms of image aestheticclassification and evaluation are to study how to make computers access the aesthetics bysimulating the visual system and aesthetic idea of human. And image aesthetic enhancementis to promote the aesthetics of images following the common rules of aesthetics. At present,there are still some problems in researches on image aesthetic classification, evaluation andenhancement, including large amount of computation in feature extraction, the incompatibilityof regular algorithms to face images and how to enhance the aesthetics of social imageseffectively.In this paper, we mainly study on the aesthetic classification, evaluation and enhancementwith social images. Our works include a method of parallel feature extraction in aestheticclassification and evaluation based on cloud computing, feature study and aestheticclassification in social image with human, aesthetic enhancement through compositionoptimization. Our works are as follows:1. To accelerate the image aesthetic process, we parallelized the feature extraction part basedon the previous work of our laboratory. And we employed the feature extraction, classificationand evaluation parts on Aliyun cloud platform to build a smart terminal application of imageaesthetic classification and evaluation.2. For those images with human, we considered the specific visual perception upon humanface, and put forward some new aesthetic features. Based on existing features, we speciallydesigned and extracted a set of face aesthetic features. With the extracted global features andsalient features, we built a classifier using SVM algorithm. The correct rates of our test on 2existing human image bases are 88.4% and 97.7%, higher than published works.3. In aesthetic enhancement, we firstly detected the subject and region dividing line. Then,following the common aesthetic rules we retargeted the subject using an improved inpaintingmethod, and then resized the regions on both sides of the dividing line using seam-carvingalgorithm. The complex details in image would be adaptively protected. The test showed theimages enhanced by our method had higher aesthetics.