Learning similarity for image retrieval with locally spatial information feedback

For content-based image retrieval techniques, systems adopt certain metric to measure the similarity between images using some visual features. In most of existing approaches, the measurements are fixed up and calculated based on the whole image. The usefulness of those systems is limited because 1. The end-users are not involved in the retrieval processes. 2. The spatial information is not taken into account and end users can not put different emphasizes on different regions in the image when retrieve. Therefore, the systems demonstrate a disability to use the locally spatial relevance feedback. This paper describes an approach towards those problems to improve the discrimination power using locally spatial information. Images are divided into 5 overlapped regions and 5 neural networks are employed to measure the similarity between corresponding regions in a query image and images in the database. The relevance feedbacks for different regions from end users are used to refine the neural networks for adjusting the measurements. Results demonstrate a good potential of the methodology.