In dealing with the explosive growth of web images, Web image annotation has become a critical research issue in recent years. Sparse feature selection plays an important role in improving the efficiency and performance of Web image annotation. In this paper, a feature selection framework is proposed with enhanced sparsity for Web image annotation. The new framework, termed as semi-supervised sparse feature selection based on l_2,1/2-matix norm with shared subspace learning (SFSLS), selects the most sparse and discriminative features by utilizing l_2,1/2-matix norm and obtains the correlation between different features via shared subspace learning. In addition, SFSLS uses graph Laplacian semi-supervised learning to exploit both labeled and unlabeled data simultaneously. An efficient iterative algorithm is designed to optimize the objective function. SFSLS method is compared to other feature selection algorithms on two Web image datasets and the results indicate it is suitable for large-scale Web image annotation.