Abstract
Distinctive and robust feature representation plays a crucial role in various multimedia applications. The descriptors invariant to rotations and textural viewpoints are able to extract discriminant features. In this paper, an improved feature descriptor named histogram of low-level visual features (HILL-VF) is proposed to extract distinctive features. In HILL-VF, fused edge maps based on gradient magnitude (FEM-GM) are obtained by using the directional derivative filters of Sobel and Scharr operators in YCbCr color space. Diffusion equation integrates the inherent edge details extracted by Sobel and Scharr edge detection operator by giving higher weights to the chroma components of the image. Moreover, phase maps based on Gabor features (PH-MGF) characterize the phase information by using Gabor filters by varying different orientations. Micro-FEM-GM and PH-MGF are generated by encoding the FEM-GM and PH-MGF into pre-defined intervals based on the selected seed values. These encoded micro-maps are then represented through a 2-D histogram. Experimental evaluations are conducted on four standard benchmarks, i.e., Coil-100k, KTH-TIPS, KTH-TIPS2-a, and -b. Experimental results indicate that we are able to increase the classification accuracy to 96.97%, 87.5%, 97%, and 93.01% on Coil-100, KTH-TIPS, KTH-TIPS2-a, and -b, respectively.