K-homogeneous nearest neighbor-driven discriminant graph coupled nonnegative matrix factorization for low-resolution image recognition

The coupled nonnegative matrix factorization method can utilize the information in high-resolution images to assist in the extraction of local semantic features of low-resolution (LR) images. However, since the supervised information is not fully considered in the existing methods, the discriminative performance of the extracted features is limited. In this paper, k-homogeneous nearest neighbor-driven discriminant graph coupled nonnegative matrix factorization (KHNNDG-CNMF) is proposed for low-resolution image recognition (LRIR). In the proposed approach, a k-homogeneous nearest neighbor-driven discriminant graph (KHNNDG) is constructed, which is a discriminant graph matrix constructed within the geometrical nearest neighbors of k homogeneous samples. In discriminant graph construction methods, the two problems of insufficient utilization of data information in local neighborhoods existing in the previous neighbor graph and the inability to reflect the real data distribution in the local neighborhood can be improved. According to the geometrical spatial distribution of samples, the KHNNDG can adaptively reflect the relationship between intra-class samples and inter-class samples, and relatively few hyperparameters are required. Therefore, the coupled nonnegative matrix factorization algorithm combined with the KHNNDG embedding regularized term can more accurately and effectively utilize the local discriminativeness of the original space to constrain the coupled features. Furthermore, we further strengthen the class separability among features by incorporating a consistent projection constraint from features to labels. The proposed algorithm performs experiments on 6 image databases. Three different LR images are involved in the experiment. Compared to the best-performing comparative method, the proposed method shows an average improvement of approximately 2.36% in recognition performance. Particularly in tasks involving extremely LR levels ( \(8\times 7\) $8\times 7$ ), the proposed method achieves an average improvement of 3.03%. Ablation experiments show that each module in the method can improve the recognition performance of LR images.