In recent years, dictionary pair learning (DPL) and deep dictionary pair learning (DDPL) have made remarkable progress in image classification tasks. However, existing DDPL methods only impose dictionary constraints on the deepest layer, ignoring different level representations. Moreover, they need to learn synthesis dictionaries for each class, and only impose inter-class constraints on the loss function, resulting in significant space consumption and loss of discriminative information. To address these issues, we introduce an innovative intra-class information guided discriminative deep dictionary pair learning (IIGD \(^{3}\) PL) method. Specifically, the IIGD \(^{3}\) PL method stacks multiple DPL layers to conduct different level representation learning. Besides, to avoid space waste, a shared synthesis dictionary is used to improve the original DPL model. At the same time, to fully utilize the category information, a discriminative constraint term guided by intra-class information is introduced to each DPL layer, which can enhance the intra-class compactness layer-by-layer. Extensive experimental outcomes indicate that the proposed method attains the pinnacle of accuracy in a majority of visual classification tasks.