Past few decades has witnessed a paradigm shift in object classification where individual feature analysis has given way to multimodal solutions. The semantic gap between different modalities such as text and image still continues to be a challenge resulting in significant amount of research and resources being devoted to the same. One crucial aspect of any multimodal task is to combine the identified features appropriately so that an optimal result can be obtained with enhanced accuracy. Although there are quite a few such feature combination techniques in literature, one can observe enough scope for improvement. In this paper, we try to address this problem of optimal feature combination for classification task using Hill Climbing. To overcome the shortcomings of Hill Climbing we propose an improvised version that increases the classification efficiency and accuracy significantly. Thorough experiments on a standard dataset using established metrics substantiate that our proposed method outperforms state-of-the-art feature combination techniques.

Identifying partially duplicated text segments among documents is an important research problem with applications in plagiarism detection and near-duplicate web page detection. We investigate the problem of local similarity search for finding partially replicated text, focusing on its parallel implementation. Our aim is to find text windows that are approximately similar in two documents, using a filter verification framework. We present various parallel approaches to the problem, of which input data partitioning along with the reduction of individual index maps was found to be most suitable. We analyzed the effect of varying similarity threshold and number of processes on speedup, and also performed cost analysis. Experimental results show that the proposed method achieves up to 13x speedup on a 24-core processor.

In this paper, we propose a novel approach which uses full processor utilization to compute a particular class of dynamic programming problems parallelly. This class includes algorithms such as Longest Common Subsequence and Needleman-Wunsch. In a dynamic programming, a larger problem is divided into smaller problems which are then solved, and the results are used to compute the final result. Each subproblem can be considered as a stage. If computations made in a stage are independent of the computations made in other stages, then these stages can be calculated in parallel. The idling of processors bottlenecks the performance of the currently existing parallel algorithms. In this paper, we are using rank convergence for computation of each stage ensuring full processor utilization. This increases the efficiency and speedup of the parallel algorithm.