Content-based image retrieval (CBIR) is vital for organizing and querying large image sets in useful ways. The general approach is to provide an image, or an image description, and find images that “look like this one.” Challenges include dealing with size and pose variation, sensitivity to relevant image features, and most of all the definition of a description that is general across a wide range of image types.

In this paper, Verma and Raman describe a particular application of local binary pattern encoding to the problem of CBIR. In basic local binary pattern encoding, each image pixel is represented by a vector composed of the signs of the differences between the center pixel and various neighborhood pixels. The matrix of these vectors at each pixel location reveals information about the local gradients in the image and provides a surprisingly good representation of both objects and texture. The authors extend this method to a center-symmetric method, where bits in the vector indicate the polarity of difference between pairs of pixels diametrically opposed around the center pixel. They label this encoding “center symmetric local binary pattern” (CSLBP). For a given image, the CSLBP at each pixel is computed for several combinations of comparison radius and angle, and these results are concatenated into a feature vector. The feature vectors are summed across the entire image, forming a compact representation of the image data.

Of course, compactness is not a virtue unless the representation also provides some discrimination. Their CBIR approach is to compute the feature vector for a test image and to find the candidates in an image set whose feature vectors are the most similar. They evaluate the use of d1, Euclidean, Chi-square, Manhattan, and Canberra vector distances; the d1 yielded the best performance. The authors implemented their own CSLBP encoding as well as several other variants of local binary patterns. In their tests, the CSLBP performed best, as measured by the most correct set of “candidate” images returned from the image sets. These results are demonstrated on sets of face images, general texture images, and brain CT slice images. If the authors had shown images that cause the method to fail, it would have aided the reader’s understanding of the method’s performance. In addition, it is vital to understand the performance on larger databases when considering the practical use of such a method, but the largest image set examined in this study consisted of only 2,800 texture images, and image subdivision was done to arrive at this size.

This paper is not only a presentation of a specific form of local binary pattern encoding for CBIR, but it also contains a good walkthrough of the basic method of local binary pattern (LBP). The language is rough in spots, but not enough to impede readability or understanding. The discussion of existing art is well done and very well referenced. This is a nice contribution to work in content-based image retrieval and is worth reading.