Photo Semantic Understanding and Retargeting by a Noise-Robust Regularized Topic Model

Retargeting aims at displaying a photo with an arbitrary aspect ratio, wherein the visually/semantically prominent objects are appropriately preserved and visual distortions can be well alleviated. Conventional retargeting models built upon perception of photos from family pre-specified communities ( e.g. , “portrait”), underlying community-specific features not learned explicitly. Thus they cannot retarget aerial photos, which contains rich variety different scales. In this work, novel framework is designed by encoding deep automatically detected Google Maps into regularized probabilistic model. Specifically, we first propose enhanced matrix factorization (MF) algorithm to calculate based on million-scale pictures, for each feature simultaneously. The MF incorporates label denoising, between-communities correlation, collaboratively. Subsequently, model called LTM that quantifies spatial layouts multiple in hidden space. To alleviate overfitting imbalanced numbers regularizer added LTM. Finally, leveraging LTM, shrink test horizontially/vertically maximize posterior probability retargted photo. Comprehensive subjective evaluations visualizations have demonstrated advantages our method. Besides, competitively consistent ground truth, according quantitative comparisons 2M photos.