A content-based image retrieval technique with tolerance via multi-page differentiate hashing and binary-tree searching multi-object buckets

Abstract

Involve manually tagging/annotating each image, and using traditional discrete data sorting techniques, such as hashing, to search images using the tags/annotations/groups of annotations. As of 2009, Flickr had 3.4 Billion Images, PhotoBucket had 7.2 Billion, Facebook had 15 Billion, and ImageShack had 20 Billion. However, none of these sites allow searching by image content and use other technologies, such as textual tags or the like. Image automatic annotation is still in its infancy. Google does allow searching in general by some kind of image-content description tagging, using some kind of limited-dictionary for image textual annotation. Although several high performance computing and data storage libraries exist, such as Hadoop and Spark, few are designed for fuzzy-with-some degree-of-similarity image non-textual content data-content retrieval. In this paper we use a multi-page hashing scheme to search images using the image itself to not only be efficient for identical images, but similar images to some degree of fuzziness and degree of similarity as well. The proposed technique uses Fourier descriptors as one representation of image objects as inputs to an evenly distributed and differentiable hashing scheme. One of the challenges in content-based retrieval schemes is the problem of overflow, usually expected in large databases. In the proposed method a Binary-Search-Tree (BST) scheme is used to decrease the search time within buckets and across pages when overflow occurs. Additionally, the proposed method allows for image retrieval based on either image object boundary contours (we call here Lambda search) or on object textures (we call here Lambda2 search) with identical or varying degrees of similarity. Benchmarking Results are presented to show the potential of the proposed method.