Tensor Feature Extraction Method Based on Matrixed Label and Inpainting

Due to the ability of tensors to maintain the potential structure of complex data and effectively describe high-dimensional data, tensor-based methods have been widely studied and applied. T-product-based representation learning methods have attracted attention because their objective functions are similar in format to vector-based methods. As t-product-based tensor representation learning methods involve block circulant matrices and Discrete Fourier transforms, there is no one-to-one correspondence between the learned reconstruction coefficients and dictionary atoms. Therefore, It is difficult to match the coefficients with labels. In order to effectively use the label information and extract discriminative features, this paper proposes a new method which treats label information as inherent information of samples and assumes that unlabeled images are contaminated with noise opposing the label information. By novelty treating labeled samples and matrixed labels as dictionaries, the proposed method performs inpainting on unlabeled samples, yielding masked matrixed labels as discriminative features for classification. The experiments prove that the proposed approach can effectively utilize label information to learn and extract discriminative features.