Modifying memorability of facial images using attention guided generative adversarial network

Humans possess the remarkable ability to vividly remember visual details, even after a single encounter. Interestingly, individuals tend to remember and forget similar things, suggesting the existence of universal internal properties within images that encode and distinguish comparable information. Research indicates that certain images are more likely to be remembered than others, and a key characteristic of such memorable images is their ability to be recalled by multiple viewers. However, creating memorable images in various domains such as visualization and photography is a challenging task. To address this challenge, the present research explores the utilization of an Attention-Guided Generative Adversarial Network (AGGAN) for modifying the visual memorability of images. Although AGGAN was not specifically designed for modifying image memorability, its capability in high-quality unpaired image-to-image translation is leveraged in this study for modifying image memorability. The objective of this research is to alter the memorability of face images by making slight attribute modifications while preserving their originality. To accomplish this, a two-step approach is adopted. First, the memorability scores for the images in the CelebA dataset are calculated by training the ResMem memorability prediction model on the 10 K US Faces dataset, which provides images with assigned memorability scores. Next, AGGAN is trained to perform image-to-image translation on unpaired images, combining the transformed image with an image from the target domain. This generalization makes AGGAN a suitable choice for the task. Experimental results on the CelebA dataset demonstrate the successful modification of image memorability using the proposed approach.