Maximal multi-channel local binary pattern with colour information for CBIR

Abstract

Content Based Image Retrieval (CBIR) focuses on retrieving images from repositories based on visual features extracted from the images. Texture and colour are one of the popularly used feature combination in CBIR. A major challenge in colour image retrieval is the characterization of features of the constituent channels and their integration. The commonly adopted methodology include extraction of features of various channels followed by their concatenation. However, the resulting image feature vector is generally of high dimensionality. To address this problem, in this paper a texture-colour descriptor is proposed integrating the multi-channel features. For texture computation, a fixed sized local intensity based descriptor, Maximal Multi-channel Local Binary Pattern (MMLBP), which integrates the multi-channel local binary information through an adder-map followed by thresholding is introduced. The histogram of the obtained patterns is used for representing the image texture. Colour information is captured by quantizing the RGB colour space and is represented with histogram. The colour-texture descriptors are further fused to characterize the images. The efficacy of the descriptor is evaluated by carrying out retrieval on benchmarked datasets for image retrieval such as Wang’s 1 K, Corel 5 K, Corel 10 K, Coloured Brodatz Texture and Zubud, using precision and recall measures as evaluation metrics. It is observed that the proposed descriptor presents improved retrieval performance over the databases under consideration and outperforms other descriptors.