<p>Fine-grained image recognition (FGIR) presents significant challenges due to the subtle distinctions between highly similar subcategories. Most existing methods, which typically focusing on a single feature granularity, often fail to capture the nuanced differences essential for accurate classification. To address this problem, we propose the Progressive Granularity-Aware Training Network (PGA-Net), a novel framework designed to enhance the discriminative power of FGIR models by incorporating multi-granularity feature recognition. Specifically, our PGA-Net utilizes a Partial Dislocation Module (PDM) to generate images of various granularities for input, and is complemented by a progressive granularity-aware training methodology that adjusts the receptive fields dynamically to integrate both coarse and fine feature details at different stages. Further refinement is achieved through a Feature Enhancement Module (FEM), which intensifies the discernibility of multi-granularity features. Additionally, PGA-Net employs a dual supervision strategy, combining cross-entropy loss and center loss, which allows it to train deep learning models effectively to capture distinct inter-class features and strong intra-class similarities, all without the need for bounding box or part annotations. This methodology allows the network to be trained end-to-end, enhancing both efficiency and accuracy. Extensive experiments and analyses on three fine-grained benchmark datasets demonstrate that the proposed PGA-Net approach can achieve new state-of-the-art.</p>