基于低秩表征学习的图像记忆性预测模型

Image Memorability Prediction Model Based on Low-Rank Representation Learning

图像记忆性预测包含两个核心问题:特征表征与预测模型。当前对图像记忆性预测的研究多聚焦于探索对其有影响的视觉因素,预测过程采用特征处理与预测相分离的方式,这使预测性能很大程度上受前期特征处理的制约,如果整个预测过程缺少整体性的学习机理,可能会产生次优的预测结果。为解决上述问题,提出了一种基于低秩表征学习的图像记忆性预测模型,将低秩表征学习和线性回归整合到一个框架下。低秩表征学习将原始的特征矩阵映射到具有低秩约束的潜在子空间中,以学习到本征稳健的特征表征;线性回归学习了一个回归系数从而建立图像特征表征和记忆性分数之间的联系。基于增广拉格朗日乘子法求解以保证模型的收敛性,大量实验结果表明本文方法的优越性。

Image memorability prediction involves two problems,feature representation and prediction model.Most of previous researches just focused on addressing the first problem by investigating the factors making an image memorable,and conducted feature fusion and regression learning in two separate steps.Results of feature fusion decide the performance of regression.Lack of using an integrated learning mechanism cannot efficiently address image memorability prediction tasks,since it may lead to sub-optimal prediction results.To solve the problem presented above,we introduce a novel image memorability prediction model based on low-rank representation learning.We seek the lowest-rank representation among all the samples by projecting the original feature matrix into a subspace spanned by a low-rank projection matrix. Meanwhile,we learn a regression coefficient to build connections between latent low-rank representations and memorability scores by linear regression.Furthermore,we develop an effective algorithm based on the augmented Lagrange multiplier method to solve our model.Extensive experiments conducted on publicly available image memorability datasets demonstrate the effectiveness of the proposed schemes.