Precision-Enhanced Image Attribute Prediction Model

Abstract

High-precision attribute prediction is a challenging issue due to the complex object and scene variations. Targeting on enhancing attribute prediction precision, we propose an Enhanced Attribute Prediction-Latent Dirichlet Allocation (EAP-LDA) model to address this issue. EAP-LDA model enhances the attribute prediction precision in two steps: classification adaptation and prediction enhancement. In classification adaptation, we transfer image low-level features to mid-level features (attributes) by the SVM classifiers, which are trained using the low-level features extracted from images. In prediction enhancement, we first exploit its advantages in extracting and analyzing the topic information between image samples and attributes by the LDA topic model. We then use a strategy to search the nearest neighbor image collection from test datasets by KNN. Finally, we evaluate the accuracy onHAT datasets and demonstrate significant improvement over the baseline algorithm.