Prototype learning is often constrained by prototype redundancy and inter-class feature confusion, limiting its interpretability and performance. This paper proposes a Decoupled Prototype Network based on Grassmann Manifold Orthogonal Projection (DPN-GOP). The method introduces a Grassmann Orthogonal Projection (GOP) module between the convolutional features and the prototype layer. By utilizing Singular Value Decomposition (SVD), it projects feature channels onto orthogonal subspaces on the Grassmann manifold to achieve effective feature decoupling. A triple orthogonal constraint loss is further designed to jointly enforce mutual exclusivity in the feature space, prototype space, and classifier weight space, thereby promoting highly discriminative representations. Moreover, a progressive three-stage optimization strategy is proposed, including geometric structure pre-training, discriminative fine-tuning, and prototype sparsity refinement. Experiments on CUB-200-2011 and Stanford Cars datasets show that DPN-GOP not only improves interpretability with more distinct prototypes, but also increases classification accuracy by 3.3% and 0.6%, respectively, while significantly reducing prototype redundancy.