홈 › Papers

Modeling Visual Memorability Assessment with Autoencoders Reveals Characteristics of Memorable Images

Background: Image memorability refers to the phenomenon where certain images are more likely to be remembered than others. It is a quantifiable and intrinsic image attribute, defined as the likelihood of being remembered upon a single exposure. Despite advances in understanding human visual perception and memory, it is unclear what features contribute to an image's memorability. To address this question, we propose a deep learning-based computational modeling approach. Methods: We modeled the subjective experience of visual memorability using an autoencoder based on VGG16 Convolutional Neural Networks (CNNs). The model was trained on images for one epoch, to simulate the single-exposure condition used in human memory tests. We investigated the relationship between memorability and reconstruction error, assessed latent space representations distinctiveness, and developed a Gated Recurrent Unit (GRU) model to predict memorability likelihood. Interpretability analysis was conducted to identify key image characteristics contributing to memorability. Results: Our results demonstrate a significant correlation between the images memorability score and autoencoder's reconstruction error, and the robust predictive performance of its latent representations. Distinctiveness in these representations correlated significantly with memorability. Additionally, certain visual characteristics, such as strong contrasts, distinctive objects, and prominent foreground elements were among the features contributing to image memorability in our model. Conclusions: Images with unique features that challenge the autoencoder's capacity are inherently more memorable. Moreover, these memorable images are distinct from others the model has encountered, and the latent space of the encoder contains features predictive of memorability.

Code (1)

Tasks

Similar Papers 제목 키워드 기반