Modular Deep Learning for advertisement image memorability: Object and text bias

Abstract

Hundreds of advertisement images are generated daily around us. They are composed of with different contents.  A few of these are only with object(s), some of these are only with text and rest is with both object(s) and text as content. Out of these some are remembered and rest of them goes out of mind. More memorable advertisement images convey their messages more conveniently to the end users. The degree or extend to which these advertisement images are remembered or forgotten is matter of concern for the product owners or sellers. In this paper we try to analyze the correlation between the image memorability and the image content at object and text level, and also try to predict the advertisement image memorability using modular deep learning. For these purposes we conducted a memory game and proposed a modular neural network using ResNet-50. Result analysis from our proposed model MResNet and ResMem revealed that images with only text as content are more memorable and text as content modifies the memorability of an image. Also from the memorability game, it was found that memorizing ability decreases when the task of memorization increases.

URN:NBN:sciencein.jist.2024.v12.740