Image Aesthetics Quality Distribution Prediction And Quality Ranking

In recent years,image aesthetics has become an important criterion for visual content curation on social media sites and media content repositories,and image aesthetics quality assessment has also become a hot topic in the field of computer vision.Previous researches on aesthetic prediction models mainly focused on aesthetic score regression or aesthetic binary classification,without considering the subjectivity of aesthetic assessment.Due to different subjective preferences,people have different aesthetic tendencies to images.A single aesthetic score or category label is not enough to characterize the differences between users' aesthetic preferences.Considering that the original aesthetic annotation is presented in the form of score histogram,the information it provides is richer and more accurate than binary label or mean score.Therefore,in this paper,we use the representation form based on aesthetic score distribution to describe the aesthetic quality of images.Through the research,we find that the distribution pattern of aesthetic score is also different when the average value of aesthetic score is in different range.Therefore,we propose a multi-channel weighted network to predict the distribution of aesthetic quality of images.The network obtains the final distribution prediction result by weighting the prediction results of different aesthetic distribution modes.We carry out the distribution prediction experiment on AVA,which is the largest aesthetic dataset at present,and compare it with the existing aesthetic quality distribution prediction algorithms.The experimental results show that our proposed algorithm obtains the optimal aesthetic quality distribution prediction results.Furthermore,we can easily transform the predicted aesthetic distribution into aesthetic score or aesthetic category,and compare it with the existing aesthetic quality classification or regression algorithms.The results show that although our network model only trains for the distributed prediction task,it gets the best results in the two classic tasks,aesthetic quality classification and aesthetic score regression.With the popularity of image aesthetic quality assessment in image retrieval,image editing and other practical applications,image comparative ranking based on image aesthetic quality is more reasonable than binary classification or score regression.However,in the current aesthetic ranking tasks,the extraction of aesthetic features is either pre-defined features or features of selflearning in deep network for ranking training.Such features cannot fully describe the aesthetic attributes of the image.Through some experiments we have proved that the extracted aesthetic features of the network that trains the aesthetic score distribution prediction have played a very good role in expressing aesthetic attributes in the tasks of aesthetic classification,regression and distribution prediction.In order to optimize the extraction of aesthetic quality features of images by the network,we designed a multi-task model that simultaneously predicts the aesthetic quality distribution of images and the relative ranking score of images.The model optimizes the extraction of image aesthetic features by using the supervision information of original image aesthetic quality distribution,and uses the two-channel network to compare and rank the input image pairs.The experimental results prove that the aesthetic features extracted by the multitask model we proposed are more effective and obtain better ranking performance.