Research Of Image Aesthetics Assessment And Cropping

Image aesthetics assessment and image cropping are two hot research directions in computer vision.Image aesthetics assessment methods aim to automatically measure whether one image is of high quality.It can be applied to image retrieval,image editing,and image recommendation.For instance,existing image retrieval algorithms mainly rank the search results based on the correlation degree to the key words and do not take into account the quality of the images.The image aesthetics assessment algorithms can remedy the deficiency.Image cropping is one common operation to improve the quality of images through cutting out the unwanted regions.Effective image cropping methods can not only help the editor save time but also help the professional crop the image.Plus,in the era of big data,automatic image aesthetics quality assessment and image cropping algorithms are of much importance.Since there are so many images uploaded to the Internet every day.Manually assessing and cropping these images would cost too much labor and time.Automatic image aesthetics assessment and image cropping algorithms can free people from these tasks.Recently,many works about aesthetics assessment have been done by international and domestic scholars.But these works are non-personalized.They assess the image aesthetics quality by assigning the image a universal label or score.These methods have similar processes.Firstly,they extract the visual features of the image,then a classifier is used to judge whether the image is of high quality.However.These methods ignore a fact that aesthetics is a subjective perception of users.Different users have different aesthetics perception.Therefore,one challenge of image aesthetics assessment is how to understand and model users' aesthetics perception and then realize personalized aesthetics assessment.In light of this,we propose one personalized image aesthetics assessment method in chapter 2.We realize image aesthetics assessment from a new perceptive by using the visual contents and the social information from social media platforms.Social information mainly includes user-image favoring information,users' contacts information and images' tags information.By mining the social information,we can model the users' aesthetics perception and then realize personalized image aesthetics assessment.We adopt two benchmark datasets for image quality assessment and collect some professional images and associated social information from Flickr.The experiments demonstrate that our method is more effective for personalized image aesthetics assessment.One image may not look beautiful overall while its visual quality can be improved by cropping operation.Existing image cropping methods mainly repurpose the classifier to perform cropping.These methods are typically multi-step frameworks.Firstly,they generate some candidate regions and then extract the features of these regions.Finally,they use a classifier or search method to find the optimal cropping region.However,these methods are slow and hard to optimize because each step needs to be trained separately.In light of this,we propose an automatic end-to-end image cropping method with a single network in chapter 3.This framework is built based on fully convolutional neural network.So it can accept the image of any size and can be optimized end-to-end on cropping performance.We test the proposed method on a public dataset released in 2017 and the experimental results demonstrate that the single network based image cropping method obtains a significant improvement on both accuracy and speed.