Diabetic Retinopathy (DR) is a significant worldwide contributor to preventable vision impairment, presenting a considerable challenge to both public health and the preservation of sight. Accurate DR grading remains challenging due to considerable variability in lesion size and morphology, indistinct boundaries, and subtle characteristics that can resemble normal retinal tissue. To tackle these challenges, we propose a novel Transformer-based framework that integrates dual-domain perception with fuzzy learning to enhance DR grading performance. Specifically, we design an Inverted Residual Fuzzy Block (IRFB) that improves lesion localization by assigning adaptive fuzzy weights across both channel and spatial domains, thereby strengthening lesion-relevant features while suppressing irrelevant information. In addition, we introduce a Fuzzy Learning-based Multi-Scale Feature Enhancement (FMFE) module, which refines multi-scale representations and reduces feature redundancy. To further advance global lesion feature extraction and contextual representation, we develop the Dual-Domain Perception Transformer (DDPT). This module leverages domain-specific self-attention to capture both spatial and frequency characteristics and applies cross-attention to fuse complementary information across domains. By jointly modeling spatial and frequency domain features, our framework achieves deeper contextual understanding and more robust representation of complex lesion structures. The experimental results demonstrate that the proposed method achieves a Quadratic Weighted Kappa (QWK) of 94.7% and an accuracy of 90.8% on the APTOS-2019 dataset. On the DDR dataset, it attains a QWK of 85.9% and an accuracy of 86.3%. These results outperform existing state-of-the-art methods, highlighting the effectiveness and robustness of the proposed approach.