【正文】 he Gabor wavelet features, the color features, etc. Specifically, a binary image is first derived by extracting feature pixels from a given image, and then a distance vector field is obtained by puting the distance vector between each pixel and its nearest feature pixel defined in the binary image. Based on the distance vector field and the FLBP parameters, the FLBP representation of the given image can be formed. In contrast to the originalLBPthat only pares a pixel with the pixels in its own neighborhood。 the FLBP can pare a pixel with the pixels in its own neighborhood as well as in other neighborhoods. The experimental results on eye detection using the BioID and FERET databases show the feasibility of our FLBP method. In particular, first, the FLBP method significantly improves upon the LBP method in terms of both eye detection rate and eye center localization accuracy. Second, we present a new feature pixel extraction method—the LBP with Relative Bias Thresholding (LRBT) method. The new LRBT method helps improve the FLBP eye detection performance when pared with other feature pixel extraction methods. Third, the FLBP method displays superior representational power and flexibility to the LBP method due to the introduction of the feature pixels as well as the FLBP parameters. Finally, in parison with some stateoftheart methods, our FLBP method achieves the highest accuracy of eye center localization.Keywords Feature Local Binary Pattern (FLBP)。 Local Binary Pattern (LBP)。 Eye detection。 Distance vector。 The LBP with Relative Bias Thresholding (LRBT)1. IntroductionThe Local Binary Patterns (LBP) method, which defines a grayscale invariant texture description by paring a center pixel with its neighbors, is a popular method for texture analysis[1],[2]and[3]. At an earlier stage for texture analysis, Wang and He[4]introduced the concept of texture unit and texture spectrum. A texture unit of a pixel is represented by eight elements, which correspond to the eight neighbors in a 33 neighborhood with three possible values: 0, 1, 2. The three values represent three possible relationships between the center pixel and its neighbors: “l(fā)ess than”, “equal to”, or “greater than”. As a result, there are 38=6561 possible texture units in total. A texture spectrum of a region is defined by the histogram of the texture units over the region. The large number of possible texture units, however, poses a putational challenge. To reduce the putational burden, Ojala et al.[1]and Gong et al.[5]proposed a method that applies two relationships: “l(fā)ess than or equal to” or “greater than” that are represented by two possible values: 0 or 1. The method thus reduces the total number of texture units from 6561 to 256, which can be represented by eight binary numbers. The two relationship version of texture units is named as Local Binary Patterns or LBP[1].The LBP method has been applied in manypattern recognitiontasks. While we are inspired by the achievement of LBP, two problems occurred to us. First, LBP only pares a pixel with the pixels in its own neighborhood. We believe that more information could be revealed if we can pare a pixel with the pixels in other neighborhoods. However arbitrarily paring a pixel with any other neighborhoods might not provide useful information. Our first problem is how to locate a pixel and a neighborhood which will provide useful information after paring them each other. Second, LBPs encodes a little information about the relationship of local texture with the features, such as edges, peaks and valleys. Our second problem is how to design a texture descriptor which encodes more information of the relationship of local texture with the features. These two problems motivated us to this study. The goal of our study is to find a new texture descriptor which can solve the two problems.We present in this paper a new Feature Local Binary Patterns (FLBP) method that encodes the information of both local texture and features. The features are broadly defined by any features which meet the requirements of specific applications, such as the edges, the intensity peaks or valleys, the Gabor wavelet features[6]and[7], the color features[8],[9],[10]and[11], as well as the LBP with Relative Bias Thresholding (LRBT) features (seeSection ). The contributions of the paper are as follows:?A new FLBP method is presented. The FLBP encodes both local and feature information. In contrast to the original LBP that only pares a pixel with the pixels in its own neighborhood, the FLBP can pare a pixel with the pixels in its own neighborhood as well as in other neighborhoods. The FLBP generalize the LBP which can be considered as a special case of the FLBP. The FLBP is expected to perform better than the LBP approach for texture description and pattern recognition.?As the FLBP method encodes both local and feature information, the performance of FLBP depends on the extraction of the feature pixels. To improve FLBP performance, we present a new feature pixel extraction method, the LBP with Relative Bias Thresholding (LRBT) method.?For the application of FLBP on eye detection, experimental results using the BioID and FERET databases show that: (i) the FLBP method significantly improves upon the LBP method in terms of both eye detection rate and eye center localization accuracy。 (ii) the new LRBT feature pixel extraction method helps improve the FLBP eye detection performance when pared with other feature pixel extraction methods。 (iii) the FLBP method displays superior representational power and flexibility to the LBP method due to the introduction of feature pixels as well as its parameters。 and (iv) in parison with the state of the art methods, the FLBP method achieves the highest accuracy of eye center localization.2. BackgroundIn recent years, theLBPmethod has been applied in many1