Discriminative Regression With Latent Label Learning for Image Classification

As one of the popular and effective supervised classification methods, linear regression is extensively used in image classification. However, the zero-one labeling matrix is too strict to be conducive to linear regression methods for learning labeling information. In addition, the linear regression focuses only on the fit of the input features to the corresponding output labels and ignores the distinctiveness between the samples. To address these two issues, this paper proposes a new method, namely, discriminative regression with latent label learning, for image classification. In contrast to the other methods, the proposed method learns labeling information in the latent label space instead of the input zero-one labeling space, doing so has the advantage that the proposed method can learn the labeling information in the data more flexibly. To guide the transform matrix to learn the discriminative information in the data, a regularization term with the idea of shortening the distance between samples within a class and lengthening the distance between samples between classes is integrated into the objective function of the proposed method. To obtain the solution of the proposed model, an iterative optimization algorithm is developed. Comprehensive experiments show that the classification performance of the proposed method outperforms the current state-of-the-art methods and deep learning methods on public image datasets with small sample sizes.