Applying a Graph-Structured Index to Product Image Search

Abstract

I have already proposed ANNG (Approximate k-Nearest Neighbor Graph) as a graphstructured access method to search for image features. However when ANNG is applied to real product image searches, two issues come up. First, when a large amount of data is stored into ANNG, some nodes on the graph are likely to have too many edges due to the characteristics of the ANNG method. Second, the existing combination search methods of image feature search and product category search are not enough to provide appropriate search results to users. To address the first issue, I propose a method to eliminate the excess edges while maintaining the connectivity of the graph. To address the second issue, I also propose a method of search using a single metric space which is a combination of both category tree-based category feature space and image feature space.