Combining Mahalanobis and Jaccard to Improve Shape Similarity Measurement in Sketch Recognition

Abstract

Mahalanobis, Jaccard and others are similarity measurements which are commonly used in sketch recognition. Attempts to improve similarity measurement can be made by manipulating formulae and reducing the testing data set used but less effort are attempted to propose algorithm. Hence, the purpose of this study is to propose a new algorithm for a better method in shape recognition. To do so, Mahalanobis and Jaccard distance measures were combined to improve the similarity measure. The pre-processing involved feature analysis, shape normalization and shape perfection and data conversion into a binary. In the new algorithm, each edge of the geometric shape was separated and measured using Jaccard distance. Shapes that passed the threshold value were measured by Mahalanobis distance. The results showed that the similarity percentage had increased from 61% to 84%, thus accrued an improved average of 21.6% difference. Having this difference, the three outcomes of this study were a combined algorithm, a new technique of separating the strokes in Jaccard, and lastly, the use of extreme vertices in Mahalanobis similarity measurement to reduce computation time.