Eigenvalue Analysis with 2D-DCT and BBP for Shape Representation and Classification

Abstract

In this work, we present eigenvalue based shape descriptor which makes use of small eigenvalue and large eigenvalue along with two dimensional Discrete Cosine Transformation (2D-DCT) for the purpose of feature extraction. The DCT based features are combined with Block based Binary Pattern (BBP) and hence propose the combined classifier model for shape representation and classification. The small eigenvalue and large eigenvalue are computed for each pixel associated with a shape, capturing the structure of a shape. It is well known fact that the 2D-DCT is capable of capturing the region information and does the energy compaction. Hence, we perform 2D-DCT on these two eigenvalue based matrices to obtain compact representation of the shape and are matched using Euclidean Distance. We have also proposed a variant of local binary pattern called blockwise binary pattern (BBP) which is found to be invariant to rotation and shift of the object. The histogram features obtained due to proposed BBP are matched using Earth Movers Distance (EMD) metric. Finally, to improve the classification accuracy, we have proposed a decision level fusion strategy which integrates 2D-DCT based features with BBP. Extensive experimental results on the publicly available shape databases namely, Kimia-99 and Kimia-216 and MPEG-7 data sets demonstrate the accuracy of the proposed method and comparative analysis exhibit that the proposed approach classifies more accurately than many baseline shape matching algorithms.