Aesthetically Relevant Image Captioning

Image aesthetic quality assessment (AQA) aims to assign numerical ratings images whilst image captioning (IAC) generate textual descriptions of the aspects images. In this paper, we study AQA and IAC together present a new method termed Aesthetically Relevant Captioning (ARIC). Based on observation that most comments an are about objects their interactions rather than aesthetics, first introduce concept Aesthetic Relevance Score (ARS) sentence have developed model automatically label with its ARS. We then use ARS design ARIC which includes weighted loss function based diverse caption selector (DACS). extensive experimental results show soundness effectiveness by demonstrating texts higher ARS’s can predict more accurately accurate, aesthetically relevant captions. Furthermore, large research database containing 510K over 5 million 350K scores, code for implementing ARIC, available at https://github.com/PengZai/ARIC