Abstract: In order to overcome the difference of features between different image domains and the gap of distribution, a learning algorithm based on co-regularized sparse concept encoding is proposed in this paper. Firstly, the distribution difference and the label consistency information of image datasets are incorporated into the sparse coding model to study the robust sparse representation of the cross-domain image. Then, the low-dimensional manifold structure is excavated from the high-dimentional image feature space to form vector set, which contructs transfer sparse coding for robust image representation. The method captures the commonality underlying of the different image dataset and realizes cross-domain transfer for image tags. The experiment shows that the method achieves more robust feature representation, and its classification performance is significantly better than other related methods.

Key words: sparse coding, manifold structure, basis learning, label relevance, common feature expression

摘要： 为克服不同图像域之间的特征“差异”，跨越分布“鸿沟”，提出了一种基于正则化迁移稀疏概念编码的跨域图像分类方法。将图像域间的分布差异性和标签相关性信息融入稀疏编码模型中，以学习跨域图像的鲁棒性稀疏表示，从高维的图像特征空间中挖掘图像低维流形结构，形成基向量集，构造跨域图像的迁移稀疏概念编码。该方法挖掘不同图像域之间的共同特征表达，实现了图像标签的跨域迁移。通过在多个图像数据库中的比较实验表明，该方法获得更为鲁棒的图像特征表达，其分类性能显著优于其他相关比较方法。

关键词: 稀疏编码, 流形结构, 基学习, 标签相关性, 共同特征表达