Higher-abstraction local binary pattern : a novel descriptor for image retrieval

Abstract

Local binary pattern (LBP) is the most commonly used feature descriptor. Many of the LBP variants are widely used in content-based image retrieval (CBIR) system because of their superior discrimination property. Retrieving an image with high precision and recall rate is a challenge. The reason being, the dynamic texture properties and large number of images in the database. The color, shape, and texture information or combination of these attributes is known as the feature of an image which is used for image retrieval in the CBIR system. LBP utilizes texture information of neighboring pixels as feature which helps in the texture classification of images. Many of the available descriptors use LBP for multiple applications such as classification, shape localization, retrieval and face recognition. If an 􀀁-bit pattern has 􀀂-state changes of bits then, to define a uniform pattern 􀀂 must be less than or equal to two. If 􀀂 is greater than two then, it will represent non-uniform patterns. The uniform pattern; an important property of image texture, together with LBP is used to obtain robust texture features for gray scale and multiresolution rotation invariant classification. It is hard to meet the requirement of “high precision and recall rate” for image retrieval. LBP has less computational complexity and better discrimination ability, but it is not possible to perform high abstraction of all the relevant features at center pixel. The term higher abstraction signifies the amount of information of surrounding neighborhood pixels for a particular center pixel. The major concern of LBP is to obtain more information of local neighborhood at center pixel with better recall and precision rate for large database of images in CBIR system. Many texture descriptors have been proposed for improving the accuracy of retrieval system, but the issue of higher abstraction is considered in HA-LBP. In LBP, a single local descriptor for a specified center pixel is used, whereas in HA-LBP, multiple nearby descriptors are employed for a given center pixel to generate higher abstraction of neighborhood pixels. HA-LBP is generic in the sense that the same coding scheme can be utilized to code the binary output image as obtained from various descriptors viz. LDP, LTP, etc. The proposed HA-LBP is efficient in retrieval of images in CBIR system. We performed experiments on three benchmark publicly available image databases i.e. COREL-1K, GHIM-10K, and Brodatz texture database. The experiments resulted in significant improvement of 5.6%, 7.47% and 2.36% in average precision rate and 3.65%, 2.81% and 2.36% in average recall rate for COREL-1K and GHIM- 10K, and Brodatz databases respectively.