Text augmentation for compressed image captioning models

Abstract

The field of compressing image captioning models to be suitable for real-Time mobile devices usage remains under-explored despite its high practical value. Recent researches showed a huge progress in this topic compressing classical state-of-the-Art models using basic models size reduction methods. However, some more sophisticated approaches that showed great results for ordinary image captioning models quality improvements are often left behind. One of such techniques is captions augmentation. It appeared to help for big uncompressed models but its influence on smaller models metrics wasn't clear. In this paper we show that along with the other image captioning models compression techniques text augmentations could help to improve quality leaving models size small enough to fit mobile devices.