A comparative study on the local-pyramid approach for Content-Based Image Retrieval

Abstract

The local-pyramid approach for image representation and feature extraction is studied for the Content-Based Image Retrieval (CBIR). Lazebnik's pyramid matching kernels and the K-means clustering is used. The SIFT descriptor is deployed for feature extraction from the images, resulting in an efficient image representation scheme and reduction of the computational complexity. Histogram intersection is used to compute the similarity between the query image and the database images. The local-pyramid approach with a 3-level pyramid and a dictionary size of 100 achieves an average precision of 86.5% in retrieving images from the benchmark database, COREL 1K, and 77.35% for that with random image database.