Aspect-Ratio-Preserving Multi-Patch Image Aesthetics Score Prediction

Abstract

Owing to the spread of social networking services (SNS), there is an increasing demand for automatically selecting, editing or generating impressive images, which raises the importance of evaluating image aesthetics. We propose the first multi-patch method for image aesthetic score prediction with the original image aspect ratios being preserved. Our method just uses images for training and does not require external information both in training as well as prediction. In an experiment using the large-scale AVA dataset containing 250,000 images, our approach outperforms other existing methods in image aesthetic score prediction, especially reducing mean squared error (MSE) of predicted aesthetic scores by 0.061 (18%) and improving the linear correlation coefficient (LCC) by 0.056 (8.9%). Noticeably, the decrease in mean absolute error (MAE) by our method for images with an unbalanced aspect ratio is at most 7.9 times larger than the decrease in MAE for images with a typical digital camera aspect ratio. This result indicates that our multi-patch method expands the range of aspect ratios with which aesthetics scores of images can be predicted accurately.