Study introduces a new development in landmark retrieval models

A new approach to landmark retrieval, an area of computer vision that identifies and matches landmark images within a database, is discussed in the International Journal of Information and Communication Technology. The new approach taken by Kun Tong and GuoXin Tan of the National Research Center of Cultural Industries at Central China Normal University in Wuhan, improves accuracy and efficiency of image retrieval systems and could help developers navigate advances in computer vision applications such as object recognition, augmented reality, and autonomous vehicle control.

Landmark retrieval models usually rely on feature descriptors to analyze and compare images. These descriptors come in two forms: global and local. Global descriptors capture the overall structure and abstract qualities of an image, while local descriptors home in on fine details such as textures and spatial arrangements. This combination offers complementary information about the image being analyzed.

However, there is a lot of redundancy, which dilutes critical information, leading to inefficient processing. Moreover, the reality of captured images means differences in viewing angle, lighting conditions, and the presence of obstructions all lead to inaccuracies.

The new model uses a texture enhancement module to emphasize the important textural features even in complex scenes. The module reconstructs feature maps to amplify surface-level patterns, ensuring that even subtle or distorted textures are highlighted. This can overcome problems that arise because of the viewing angle or poor lighting. The model also uses a feature fusion module that integrates the global and local descriptors to eliminate redundancies in the data. By prioritizing relevant details and discarding superfluous information, the model streamlines the analysis to improve computational efficiency.

Tong and Tan have carried out extensive tests on benchmark datasets, including the Revisited Oxford and Paris datasets, and show their approach to be very effective and efficient at identifying landmarks.