Enhanced Image Retrieval Through the Integration of Dense Global Descriptors and Structural Masks

Similar image retrieval identifies and ranks images from a database based on visual characteristics such as color, texture, and shape to match a given sample image, providing results with the closest resemblance. Hence, it's crucial to have a thorough understanding of the content in images. In this paper, we propose a method to utilize structural information based on an image segmentation model. Significantly, the structural information can provide novel insights into aspects of images that conventional dense global descriptors may overlook, as it can clarify the shapes of objects and their backgrounds, offering a complementary perspective. Specifically, our method combines structural information with global descriptors, allowing both the detailed shapes and broader features of images to be captured in a manner that users can control. In the experiments, we evaluate the extent to which this integrated approach enhances the performance of similarity search tasks.