<p>Unsupervised cross-modal hashing, due to its independence from heavy label information, is more convenient for application to other fields. In recent years, this area has gained widespread attention and achieved great success. However, existing unsupervised cross-modal hashing methods still face some issues, such as simple fusion after feature extraction, the use of a single similarity measure to express data relationships, and guiding hash code learning through a single affinity matrix. To address these problems, we propose a new method called Dual-Matrix Guided Reconstruction Hashing for Unsupervised Cross-Modal Retrieval. We construct an effective matrix from the extracted raw semantic information to guide the generation of reconstructed hash codes for images and texts. Simultaneously, we construct another matrix for the extracted image and text features, guiding the generation of reconstructed hash codes using graph convolution, thus directing hash code learning through dual matrices. In evaluations on three standard datasets, our method achieved an average improvement of approximately 1.3% in MAP@5000 and 1.5% in MAP@50, particularly showing significant performance gains with shorter hash codes.</p>