Gabor-Based Novel Local, Shape and Color Features for Image Classification

This paper introduces several novel Gabor-based local, shape and color features for image classification. First, a new Gabor-HOG (GHOG) descriptor is proposed for image feature extraction by concatenating the Histograms of Oriented Gradients (HOG) of all the local Gabor filtered images. The GHOG descriptor is then further assessed in six different color spaces to measure classification performance. Finally, a novel Fused Color GHOG (FC-GHOG) feature is presented by integrating the PCA features of the six color GHOG descriptors that performs well on different object and scene image categories. The Enhanced Fisher Model (EFM) is applied for discriminatory feature extraction and the nearest neighbor classification rule is used for image classification. The robustness of the proposed GHOG and FC-GHOG feature vectors is evaluated using two grand challenge datasets, namely the Caltech 256 dataset and the MIT Scene dataset.