Image Aesthetic Quality Assessment Based On Convolutional Neaural Networks And Composition Rules

In recent years,with the increasing amounts of photos and social media platforms,people pay more and more attention on the aesthetic quality of images.People would like to select attractive images efficiently and accurately when browsing and saving images.When sharing and displaying photos,people prefer to select attractive pictures.Therefore,researchers focus on the reasonable and accurate judgements of image aesthetic quality.The image aesthetic assessment aims to model the image aesthetic attributes to enpower computers the ability of automatically judging the image aesthetic quality through simulating the process of human beauty-appreciation.Existing methods mainly include handcrafted feature based methods and deep learning based methods.Due to the complexity of image content and the subjectivity of aesthetics,it is difficult to construct aesthetic features accurately and comprehensively.Though the deep learning method can extract image aesthetic characteristics automatically,resizing the image to adapt the input of the network might cause the loss of many details.It could destroy the original aesthetic attributes in image.Therefore,the aesthetic quality evaluation methods still has great potential to be developed.An image aesthetic assessment method based on the attribute graph is proposed.In view of the difficulties of constructing the aesthetic features,an image composition feature is designed as the edges of the attribute graph.The feature is consisted of the distance,the angle,the overlapping area between objects,combining with the object's position information in the whole image.The low-level features of objects and the overall image's low-level features are used as the node features of the attribute graph.The combination of the nodes and the edges is used to construct an attribute graph to train a SVM classifier,and evaluate the aesthetic quality of the image.Experiments show that the method can effectively categorize the aesthetic quality of images.An image aesthetic evaluation method based on the double-column convolutional neural network with semantic information is proposed.To resolve the problem of missing aesthetic details raised by resizing image,resized images and random crop image patches are used as the inputs of two columns in the network respectively,which extracts the aesthetic characteristics of image in a global view and a local view.At the same time,a parallel convolution structure is used in the global view to eliminate the influence to the aesthetic quality caused by the semantic information.The combination of features extracted from the global view and the local view is used to train the model to assess the aesthetic quality of image.The network uses the end-to-end structure to realize the aesthetic feature extraction and the aesthetic classification automatically.The experimental results show that the method has many advantages in the aesthetic classification task.An image aesthetic assessment method based on region details and composition is proposed.Aiming at the matter of inadequate access to the aesthetic details using random cropping image in the double-column network,different image cropping strategies are used to obtain image patches,and extract the aesthetic features of the corresponding area using a multi-column convolution network.At the same time,the convolutional network is used to extract image composition features,thereby describing the structure of image patches.The combination of regional aesthetic details and composition features is fed to fully connected layers to map the features to image aesthetic quality.Finally,the experimental results show that the method has a good performance in the aesthetic classification task.