Local triplet pattern for content-based image retrieval

Abstract

An image feature named Local Triplet Pattern (LTP) is proposed for image retrieval applications. The LTP feature of an image is a histogram which contains spatial information among neighboring pixels in the image. An LTP level is extracted from each 3×3 pixel block. The color levels of the eight surrounding pixels are compared with the color level of the center pixel. The comparison returns one of the triplet codes: 0, 1, or 2 to represent the three conditions: the color level of a neighboring pixel is smaller than, equal to, or larger than the color level of the center pixel. The eight triplet codes from the eight surrounding pixels are then transformed to an LTP level. We also consider extracting the LTP from a quantized color space and at different pattern length according to the application needs. Experimental results show that our proposed LTP histogram consistently outperforms other histograms with spatial information on both the texture and generic image datasets. © 2009 Springer Berlin Heidelberg.