Aesthetic Multi-attributes Captioning Network for Photos

In recent years, image aesthetic quality assessment has become increasingly popular. In addition to numerical assessment, aesthetic captioning has been proposed to capture the overall aesthetic impression of an image. To further advance this field, we address a task of aesthetic attribute assessment, which is the aesthetic multi-attributes captioning. Labeling the comments of aesthetic attributes is a non-trivial task, which limits the size of available datasets. We construct a novel DPChallenge Multi-Attributes Captions Dataset (DPC-MACD) dataset by a semi-automatic way. We propose two novel aesthetic multi-attributes captioning networks, which are the Bottom-Up and Top-Down Attention Network (BUTDAN) and Object-Semantics Aligned Pretrained Network (OSAPN). The experimental results show that our method can predict the comments, which are more closely aligned to aesthetic topics than those produced by the previous models. Through the evaluation criteria of image captioning, the specially designed model outperforms other methods.