DoG Based Local Ternary Pattern for Image Retrieval

Abstract

In this paper, a novel pattern-based feature is proposed for the purpose of image retrieval. It extracts edge information of images using difference of Gaussian (DoG) filter. The choice of DoG filter depends upon the quality of available image and the information content, i.e, the value of blurriness (standard deviation) can be controlled and hence the edges can be highlighted more accurately. From the filtered image, texture statistics are captured using local ternary pattern (LTP) extracted from four different orientation images to form DoGLTP. The different orientations used in this work enable us to capture angle information usually ignored by many pattern-based features. In the proposed feature, color information is included through Hue and Saturation histograms. This makes the proposed feature more robust and the retrieval results highlight its superior performance in comparison to existing techniques on two popular databases: Corel 1000 and Corel 10k.