Shape-based recognition using combined Jaccard and Mahalanobis measurement / Noor Aznimah Abdul Aziz

Numerous well-established shape recognition approaches for handling variance of image transformations and strokes variations in free-hand digital sketching environment but none has satisfactorily deal with object features yet. Complexity in the computation measure, expensive in computation time, requires large training data set and extensive training is not practical, especially in sketching recognition. In contrast, methods for similarity measurement such as Jaccard distance, Mahalanobis distance and others are commonly used in recognition tasks offer a simple computation, not require a large training data set and can handle variances of image transformations and strokes variations. Therefore, we proposed a shape recognition algorithm using similarity measurement combining Jaccard and Mahalanobis distance is used to measure the similarity between geometry shape sketches. Two major pre processing procedures involved feature extraction and edges perfection were performed for shape normalization and beautification. The new combined algorithm also implements edges separation and masking technique to improve similarity measurement and reduce the amount of testing data set used. Results show that the combination of Jaccard and Mahalanobis distance increase similarity percentages from 18% to 66%, thus accrued an improvement of 48% differences. Having this difference, the two major contributions made in this study are first a combined algorithm and a new technique of separating edges in Jaccard and the use of extreme vertices in Mahalanobis distance. This proposed combined algorithms and techniques can achieve high performance in shape similarity measurement recognition and also the masking technique in Mahalanobis distance measurement can reduce the amount of data analysis.