Abstract: In recent years, the number of pictures related to weakly supervised user-provided tags has increased
dramatically. The user-provided tags are noisy, inadequate and subjective in nature. The problem of the social image
understanding is focused by this proposed system, i.e., tag refinement, tag assignment, and image retrieval. The proposed
system presents a novel weakly supervised deep matrix factorization algorithm different from past work to uncover the
latent image representations and tag representations by collaboratively exploring the weakly supervised tagging data, the
visual structure, and the semantic structure embedded within the latent topological space. To learn a semantic subspace
without over-fitting the incomplete, noisy or subjective tags the semantic and visual structures are jointly incorporated. A
sparse model is imposed on the transformation matrix of the first layer to remove the noisy or redundant visual options,
in the deep architecture. On the tasks of image understanding like image tag refinement, assignment, and retrieval the
extensive experiments conducted on real world social image databases. The effectiveness of the proposed method is
demonstrated by the achieved encouraging results.
Keywords—User Provided Tags, Image Tag Refinement, Image Tag Assignment, Image Tag Retrieval, Social Image
Understanding.