DCGMDL: Deep Conditional Generative Model for Dictionary Learning in Supervised Classification

In this paper, we propose the DCGMDL model, which utilizes a feed-forward neural network to exploit the inherent nonlinearity of data for dictionary learning. Unlike traditional dictionary learning methods that focus on learning linear dictionaries directly from nonlinear data in the original space, the DCGMDL approach introduces a novel framework by learning a linear dictionary in a transformed latent space. This design enables the model to effectively capture and utilize the nonlinear characteristics of input samples, resulting in a more robust and flexible approach to dictionary learning. To validate the effectiveness of the proposed method, we conducted experiments on image classification tasks using two benchmark datasets: Scenes-15 and ImageNet10. The experimental results demonstrate that the DCGMDL model achieves superior performance by leveraging the nonlinear properties of the data, highlighting its potential to overcome the limitations of conventional dictionary learning techniques.