Sensitivity based image filtering for multi-hashing in large scale image retrieval problems
- PDF / 3,532,158 Bytes
- 18 Pages / 595.276 x 790.866 pts Page_size
- 14 Downloads / 206 Views
ORIGINAL ARTICLE
Sensitivity based image filtering for multi-hashing in large scale image retrieval problems Wing W. Y. Ng1 • Jinchen Li2 • Shaoyong Feng1 • Daniel S. Yeung1 Patrick P. K. Chan1
•
Received: 10 February 2015 / Accepted: 14 July 2015 Ó Springer-Verlag Berlin Heidelberg 2015
Abstract Hashing is an effective method to retrieve similar images from a large scale database. However, a single hash table requires searching an exponentially increasing number of hash buckets with large Hamming distance for a better recall rate which is time consuming. The union of results from multiple hash tables (multihashing) yields a high recall but low precision rate with exact hash code matching. Methods using image filtering to reduce dissimilar images rely on Hamming distance or hash code difference between query and candidate images. However, they treat all hash buckets to be equally important which is generally not true. Different buckets may return different number of images and yield different importance to the hashing results. We propose two descriptors, bucket sensitivity measure and location sensitivity measure, to score both the hash bucket and the candidate images that it contains using a location-based sensitivity measure. A radial basis function neural network (RBFNN) is trained to filter dissimilar images based on the Hamming distance, hash code difference, and the two proposed descriptors. Since the Hamming distance and the hash code difference are readily computed by all hashingbased image retrieval methods, and both the RBFNN and the two proposed sensitivity-based descriptors are computed offline when hash tables become available, the & Jinchen Li [email protected] Wing W. Y. Ng [email protected] 1
School of Computing Science and Engineering, South China University of Technology, Guangzhou, China
2
College of Medical Information Engineering, Guangdong Pharmaceutical University, Guangzhou, China
proposed sensitivity based image filtering method is efficient for a large scale image retrieval. Experimental results using four large scale databases show that the proposed method improves precision at the expense of a small drop in the recall rate for both data-dependent and data-independent multi-hashing methods as well as multi-hashing combining both types. Keywords CBIR
Multi-Hashing RBFNN Image filtering
1 Introduction An efficient similarity measure between images in a large scale database is a key to the success of content-based image retrieval (CBIR) [1–3]. For a given query image E, the task is to find candidate images (Is) which either have similarity measures, sim(E, I),larger than a given threshold, or rank among the top k images with the largest sim(E, I). The similarity measure could be either Euclidean based or a semantically learnt one [4]. Obviously it is both inefficient and unnecessary to compute similarity measures between the query and all images in a large scale database. Two major approaches have been proposed to find a subset which is much smaller than the whole database and
Data Loading...
