An Universal Image Attractiveness Ranking Framework
share
We propose a benchmark framework to rank image attractiveness using a novel model trained with a large set of side-by-side multi-labeled image pairs. We use an efficient way to collect a large and diverse set of image pairs and directly rate each pair' relative attractiveness. The judges only need to provide relative ranking between two images without the need to directly assign an absolute score. We investigate a deep attractiveness rank net (DARN), a combination of deep convolutional neural network and rank net, to generate an attractiveness score mean and variance for each image after training on side-by-side rated image pairs. The deep neural network jointly learns a mapping from each DNN feature to a score mean and variance, and the underlying criteria the judges use to label each image pair. The score can be used as a useful feature to rank image attractiveness. We show reasonable prediction error of the model and observe significant image quality improvement when using this feature in a real commercial search engine.
READ FULL TEXT
