A sparsity constrained low-rank matrix completion approach for image tag refinement

Mengdi Liu, Lei Chen, Wencai Ye, Mengqiang Xu

Abstract: In recent years, the vast images with user-provided tags are easily available on the photo-sharing platform, which can greatly promote image retrieval and management. However, these tags often are incomplete and noisy, impeding the tag related image applications. To address this challenge, a Sparsity Constrained Low-Rank Matrix Completion (SCLRMC) model is proposed for simultaneously completing and de-noising the image tags. Compared with the previous low-rank matrix completion model, the SCLRMC model integrates with the inherent sparsity property of image tags. Furthermore, a Linearized Alternating Direction Method of Multipliers (LADMM) based on Proximal Forward Backward Splitting (PFBS) algorithm is employed to solve the proposed SCLRMC model. Experimental results on the benchmark datasets well verify the effectiveness and efficiency of our proposed SCLRMC model.