Abstract: In the field of fine-grained image retrieval, existing research mainly focuses on using deep networks to achieve discriminative feature extraction and precise localization, ignoring the importance of shallow feature information and unable to eliminate complex noise interference in the background, which limits the improvement of retrieval performance. Therefore, a Fine-grained Deep Hashing image retrieval method based on Multi-level Feature Extraction (FDH-MFE) is proposed, which mainly focuses on the correlation between features at different levels and enhances the ability to extract local features. Firstly, a feature extraction module is proposed to extract fine-grained features from different stages of the network and reveal their potential long-range dependencies through graph neural networks, providing more comprehensive and refined feature representations for subsequent stages. Secondly, a proxy loss algorithm is designed to make the distribution of hash codes more uniform, thereby enhancing the discriminative ability of fine-grained features. Finally, by designing a background suppression algorithm and combining it with ternary loss, the model's ability to fit global distributions is enhanced, making the proposed method perform well in fine-grained image retrieval tasks. The experimental results show that the average retrieval accuracy of this method on four public data sets is improved by 15.03%, 10.94%, 9.98% and 9.78% respectively compared with the sub advanced method.