| A new era of information explosion is coming with the widespread of computer network and multimedia technology,a tremendous number of images are generated.People's lives are full of various picture information,some of which make people impressive,while some gradually fade out of people's mind.Recent years,a lot of work have done to explain this phenomenon,finding out that this phenomenon is caused by image memorability,which is picture's inherent attribute and intends to describe how much can an image be remembered.At present,there are two main aspect of study on image memorability,one is to study on the factors that affect image memorability,and another focus on image memorability prediction.Although multiple types of features can be extracted to encode image memorability from different aspects,these features are generally used through simple concatenation without further refinement.Such an approach may not work well in capturing the intrinsic representation of images and easily cause information redundancy.Additionally,many works neglect the innovation of the prediction models.This article pays more attention on the prediction of image memorability and proposed two models for image memorability prediction.First,we proposed a general framework called Joint Low-rank and Sparse Principal Feature Representation for Image Memorability Prediction,which aims to achieve the lowest-rank discriminative feature representation and learn a sparse coefficient matrix to connect the feature and image memorability scores.Then for improving robustness of image memorability prediction,we proposed a method called Low-rank-sparse Representation with Robust Relationship Inference for Image Memorability Prediction.The core of the method focuses on exploring the relationship structure between image samples based on the effective feature representation.Extensive experiments were conducted on two popular public datasets in this filed and proved the superiority of our models. |
PDF Full Text Request