Learning descriptive visual representation by semantic regularized matrix factorization

摘要: This paper presents a novel semantic regularized matrix factorization method for learning descriptive visual bag-of-words (BOW) representation. Although very influential in image classification, the traditional BOW representation has one distinct drawback. That is, efficiency purposes, this is often generated by directly clustering low-level feature vectors extracted from local keypoints or regions, without considering high-level semantics of images. In other words, still suffers gap and may lead to significant performance degradation more challenging tasks (e.g., classification community-contributed images with large intra-class variations). To overcome drawback, we develop adding Laplacian regularization defined tags (easy access although noisy) into factorization. Experimental results on two benchmark datasets show promising proposed method.