PROJECT TITLE :

Enhanced Discrete Multi-modal Hashing More Constraints yet Less Time to Learn

ABSTRACT:

As a result of the meteoric rise in the amount of multimedia data, multi-modal hashing, a potentially useful method that could make cross-view retrieval more scalable, is garnering an increasing amount of interest. The majority of the existing multi-modal hashing methods, on the other hand, either divide the learning process into two unnaturally distinct stages or treat the discrete optimization problem as if it were a continuous one, both of which result in suboptimal results. Neither of these approaches is particularly natural. Recently, a few discrete multi-modal hashing methods that attempt to address such issues have emerged; however, these methods continue to ignore a number of significant discrete constraints (such as the balance and decorrelation of hash bits). In this paper, we propose a novel method called "Enhanced Discrete Multi-modal Hashing (EDMH)," which learns binary codes and hashing functions simultaneously from the pairwise similarity matrix of data, while adhering to the aforementioned discrete constraints. This allows us to circumvent the limitations that were previously mentioned. In spite of the fact that the EDMH model appears to be a great deal more complicated than the other models for multi-modal hashing, we are actually capable of developing a quick iterative learning algorithm for it. This is due to the fact that the subproblems of its optimization all have closed-form solutions once a couple of auxiliary variables are brought into the equation. The results of our experiments on three real-world datasets have shown that the discrete constraints that were previously ignored can be useful, and they have also shown that EDMH not only performs significantly better than state-of-the-art competitors in terms of several retrieval metrics, but it also runs significantly faster than the majority of those competitors.