丰富图像标签的正则化非负矩阵分解方法

目的 随着Web2.0技术的进步，以用户生成内容为中心的社交网站蓬勃发展，也使得基于图像标签的图像检索技术越来越重要。但是，由于用户标注时的随意性和个性化，导致用户提交的图像标签不够完备，降低了图像检索的准确性。方法 针对这一问题，提出一种正则化的非负矩阵分解方法来丰富图像欠完备的标签，提高图像标签的完备性。利用非负矩阵分解的方法将原始的标签-图像矩阵投影到潜在的低秩空间里消除噪声，同时利用图像的类内视觉离散度作为正则化项提高消除噪声、丰富标签的效果。结果 利用从社交网站Flickr上下载的大量社交图像进行对比实验，验证了本文方法对丰富图像标签的有效性。通过对比目前流行的优化算法，本文算法获得较高的性能提升，算法平均准确度提高了12.3%。结论 将图像类内视觉离散度作为正则化项的非负矩阵分解算法，能较好地丰富社交图像的标签，解决网络图像标签的欠完备问题。

Regularized non-negative matrix factorization method for image tag enrichment

Objective With the development of Web 2.0, social websites centered on user-generated content are arising. Therefore, tag-based image retrieval becomes more and more important. However, the image tags that users upload are incomplete because users label images freely and arbitrarily and thus decrease the performance of image retrieval. Method To solve the problem of image tag incompletion, this paper proposes an algorithm based on regularized non-negative matrix factorization to enrich the tags of social images and make these tags complete. This proposed algorithm casts the original tag-image matrix to a latent low-rank space and discovers the correlations between tags with the matrix factorization technique. The relationships among tags are utilized to enrich tags for social images. Meanwhile, the overall visual diversity as a regularization term is utilized to restrict the impact of content-irrelevant tags and enrich image tags. Result This paper constructs comparison experiments on images downloaded from sharing website Flickr. Accuracy is used to evaluate these comparison experiments. These experiments demonstrate the effectiveness of our proposed algorithm for enriching image tags. Compared with state-of-the-art approaches, our approach could improve average accuracy by 12.3%. Conclusion This paper proposes a regularized non-negative matrix factorization framework with overall visual diversity as the regularization term and enriches the tags of images effectively. Our proposed algorithm can solve the problem of incomplete tags.