Tag localization with spatial correlations and joint group sparsity

摘要: Nowadays numerous social images have been emerging on the Web. How to precisely label these is critical image retrieval. However, traditional image-level tagging methods may become less effective because global matching approaches can hardly cope with diversity and arbitrariness of Web content. This raises an urgent need for fine-grained schemes. In this work, we study how establish mapping between tags regions, i.e. localize so as better depict index content images. We propose spatial group sparse coding (SGSC) by extending robust encoding ability correlations among training regions. present in a two-dimensional space design group-specific kernels produce more interpretable regularizer. Further joint version SGSC model which able simultaneously encode intrinsically related regions within test image. An algorithm developed optimize objective function Joint SGSC. The tag localization task conducted propagating from sparsely selected groups target according reconstruction coefficients. Extensive experiments three public datasets illustrate that our proposed models achieve great performance improvements over state-of-the-art method task.