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 which is time consuming. The union of results from multiple hash tables (multi-hashing) yields a high recall but low precision 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 not true in many cases. Moreover, a large hash bucket may return many more dissimilar images, especially for those images located close to the bucket boundary. Unfortunately, it is difficult to calculate the exact bucket size owing to its high-dimensional irregular shape. We propose two descriptors to score the hash bucket size and image location using a location based sensitivity measure. A neural network is trained to filter dissimilar images based on these two descriptors, hamming distance and code difference. Both the neural network and the two new descriptors could be computed offline when hash tables are available. Hence the proposed Sensitivity based Image Filtering method (SIF) is efficient during retrieval. Experimental results using four large scale databases show that the proposed method improves precision and recall for both data-dependent and data-independent multi-hashing methods as well as multi-hashing combining both types.