Coslets: A Novel Approach to Explore Object Taxonomy in Compressed DCT Domain for Large Image Datasets

The main idea of this paper is to exploit our earlier work of image segmentation [11] and to propose a novel transform technique known as Coslets which is derived by applying 1D wavelet in DCT domain to categorize objects in large multiclass image datasets. Firstly, k-means clustering is applied to an image in complex hybrid color space and obtained multiple disjoint regions based on color homogeneity of pixels. Later, DCT brings out low frequency components expressing image’s visual features and further wavelets decomposes these coefficients into multi-resolution sub bands giving an advantage of spectral analysis to develop robust and geometrically invariant structural object visual features. A set of observed data (i.e. transformed coefficients) is mapped onto a lower dimensional feature space with a transformation matrix using PCA. Finally, different distance measure techniques are used for classification to obtain an average correctness rate for object categorization. We demonstrated our methodology of the proposed work on two very challenging datasets and obtained leading classification rates in comparison with several benchmarking techniques explored in literature.