Integrity-Preserving Image Aesthetic Assessment

Abstract

Image aesthetic assessment is a challenging problem in the field of computer vision. Recently, the input size of images is often limited by the network of aesthetic problems. The methods of cropping, wrapping and padding unify images to the same size, which will destroy the aesthetic quality of the images and affect their aesthetic rating labels. In this paper, we present an end-to-end deep Multi-Task Spatial Pyramid Pooling Fully Convolutional Neural NasNet (MTP-NasNet) method for image aesthetic assessment that can directly manipulate the original size of the image without destroying its beauty. Our method is developed based on Fully Convolutional Network (FCN) and Spatial Pyramid Pooling (SPP). In addition, existing studies regards aesthetic assessment as a two-category task, a distribution predicting task or a style predicting task, but ignore the correlation between these tasks. To address this issue, we adopt the multi-task learning method that fuses two-category task, style task and score distribution task. Moreover, this paper also explores the reference of information such as variance in the score distribution for image reliability. Our experiment results show that our approach has significant performance on the large-scale aesthetic assessment datasets (AVA [1]), and demonstrate the importance of multi-task learning and size preserving. Our study provides a powerful tool for image aesthetic assessment, which can be applied to photography and image optimization field.