Research On Image Aesthetics Quality Assessment Based On Deep Neural Network

Image aesthetic quality assessment aims to learn a model to automatically classify the image into high quality or low quality.The key task is to find a suitable image representation to establish the relationship between image content and aesthetic quality.Because of its potential application value in the fields of pattern recognition and computer vision,image aesthetic quality assessment has gradually become a hot research topic and attracted increasing attention.Not only can image aesthetic assessment play an important role in image search,image filtering,image editing,but also it has broad application prospects in poster design,manuscript design,costume design,architectural design,etc.Therefore,it has become an urgent problem to design an effective algorithm to improve the performance of image aesthetic quality assessment.Our first work is the research on image aesthetic quality classification based on users' social behavior.The key task of aesthetic assessment is to extract appropriate aesthetic features,which advance stem from designing handcrafted features to automatically learning features from the deep neural.Feature extraction has witnessed the improvement of the performance of aesthetic assessment.However,the existing methods were purely based on the visual content of images,without considering the human cognition of images.Nowadays,with the development of the internet and the emergence of many social media platforms,we can get a large number of social images with user behavior information.Social psychology research shows that human cognitive and behavior are interrelated,so we propose to obtain users' cognitive information from their social behaviors,and further use this information to improve the performance of image aesthetic assessment.First,in order to decrease the uncertainty of users' social behaviors,we cluster different types of original social behaviors and use social distribution on different clusters to represent each social image.Then,we use the idea of transfer learning to train a social behavior detector with social images and apply it to extract the cognitive features of common web images.Finally,image visual content and user cognitive information are fused to improve the performance of image aesthetic assessment.Experiments on two aesthetic assessment benchmark datasets demonstrate the effectiveness of our approach.Our second work is the label distribution research on image aesthetic assessment based on a semantic hybrid network.In some early works,images were often assigned a single label,namely high quality or low quality,or an aesthetic score,and image aesthetic quality assessment was always classified as a binary or regression problem.However,as a kind of spiritual and cultural activity,different people may have a different aesthetic perception of the same image.Therefore,the binary aesthetic label cannot well measure the aesthetic preference of all people.Different from the existing aesthetic classification researches,our work is based on label distribution learning,which quantifies image aesthetics into a distribution vector on several aesthetic levels.Our framework is based on fully convolutional network and can preserve the original image size without converting it to a fixed size,and avoid the risk of losing the inherent aesthetics of the image,so as to improve the performance of the prediction task of image aesthetic distribution.In addition,considering that aesthetic perception needs to accompany by the human semantic understanding of images,we propose to use the semantic information of images to assist aesthetic assessment task.We propose a semantic awareness hybrid network based on label distribution learning to achieve these functions.Experiments on two aesthetic assessment benchmark datasets demonstrate the effectiveness of our approach and the usefulness of preserving the original image size and using the semantic information of the image.