An investigation of combining gradient descriptor and diverse classifiers to improve object taxonomy in very large image dataset

Abstract

Assigning a label pertaining to an image belonging to its category is defined as object taxonomy. In this paper, we propose a transform based descriptor which effectively extracts intensity gradients defining edge directions from segmented regions. Feature vectors comprising color, shape and texture information are obtained in compressed and de-correlated space. Firstly, Fuzzy c-means clustering is applied to an image in complex hybrid color space to obtain clusters based on color homogeneity of pixels. Further, HOG is employed on these clusters to extract discriminative features detecting local object appearance which is characterized with fine scale gradients at different orientation bins. To increase numerical stability, the obtained features are mapped onto local dimension feature space using PCA. For subsequent classification, diverse similarity measures and Neural networks are used to obtain an average correctness rate resulting in highly discriminative image classification. We demonstrated our proposed work on Caltech-101 and Caltech-256 datasets and obtained leading classification rates in comparison with several benchmarking techniques explored in literature.