Abstract
In this paper, a new scheme of feature extraction named as sparse Weber-oriented visual features is proposed by the integration of sparse and dense data representations. Image orientation, magnitude and pixel intensity are aggregated in horizontal direction to obtain the feature vector that provides more discriminative information of an image objects. Instead of merely considering individual pixel intensities being highly susceptible to image local variations, orientation is calculated using horizontal and vertical local patches along current pixel. The magnitude component is computed using dense chromatic data representation by second-order symmetric kernel function. Pixels randomly selected from sparse data are binarized with winner-take-all principle to generate feature vector directly from pixel intensities. Experiments using KNN classifier resulted in object recognition accuracy of 94.43%, 96.21% and 71.8% on three standard available colored image datasets that are COIL-100, ALOI and PVOC 2007, respectively. Results evaluation depicts the performance of proposed method as compared to state-of-the-art object recognition schemes.