Existing image retrieval methods often suffer from the loss of structural information during local feature extraction and the high computational cost associated with local feature re-ranking. To address these issues, an image retrieval method based on self-similarity embedding and global feature re-ranking was proposed. A self-similarity embedding network was introduced to capture the internal structure of images and compress it into a dense self-similarity feature map. A self-similarity embedded feature map was generated by fusing the self-similarity feature map with the initial image feature map. The generated map could represent both the visual and the structural information of the image, thereby achieving finer-grained retrieval results. A global feature re-ranking method was proposed by drawing inspiration from query expansion and database enhancement techniques. Based on the initial ranking results, the features of the top-ranked similar images for each image were extracted and the initial features were updated using a linear summation approach. This process highlighted the common features of images with the same content and increased the inter-class distance, thereby reducing the number of false positives. In experiments, the proposed self-similarity embedding and re-ranking methods were evaluated using the mean average precision (mAP) as the evaluation metric. The results demonstrated that the proposed method outperformed existing approaches on the ROxford5K and RParis6K datasets.