The Research And Application Of Image Feature Extraction Based LBP Methodology

Feature extraction is an important concept in Computer Vision and Digital Image Processing. Image features, characterizing the basic properties of an image, can be seen as a visual expression of the image. Generally, people judge the quality of feature extraction mainly from four aspects:distinction, reliability, independence and data. The common image features include color features, texture features, shape features and spatial relation features.Texture is a fundamental property of the object surface. It reflects the structural organization of the surface arrangement about the important information and the relationship with its surrounding environment. The purposes of texture extraction are that feature dimensions are as small as possible, identification capability is strong, have a good stability, time and space complexity is low and can apply widely. There are four analysis methods:1. Statistical analysis method. It adopts together statistics from the image area to extract texture. The method describes the direction, uniformity and other basic information through gray value distribution of the image. It includes Self-function method. Tamura texture feature. Gray level co-occurrence matrix. Gray level gradient co-occurrence matrix and so on.2. Structure analysis method. This method assumes that all the complex textures can be formed by the simple textures in some rules. But it rely on the regularity od the texture too much. General methods include syntactic texture description and mathematical morphology method, and etc.3. Model analysis method. It models on the texture, so that the texture extraction problem is transformed into a parameter estimation problem. The most common model analyses are Markov (MRF) model, autoregressive texture (SAR) model and Rractal law, and etc.4. Spectrum analysis method. This method is based on the muti-scale analysis. They use some kind of filter to convert the image texture to the transform domain, and then use some rules to extract the texture. The most common methods are wavelet transform and Gabor transform.Local Binary Pattern (LBP) is used to describe the image texture features, proposed by Professor Ojala T of Oulun University in 1996. It can be classified into the statistical analysis method. LBP has many advantages:1. Easy to understand. Just compare the center pixel and the neighborhood pixels, and then calculate the binary mode, that is LBP,very understandable.2. Low computational complexity. Compared to other texture algorithms, LBP has low computational complexity because it can be obtained by scanning the image only once.3. Identify ability well. LBP can express a number of subtle features in the image. For example, bright spots, dark spots, edges and so on.Ojala improved the original LBP later, proposed a multi-scale LBP, uniform LBP, rotation invariant LBP. From that, it has translational invariance, gray monotone invariance, rotation invariance and other properties. Furthermore Uniform LBP can reduce the dimension of LBP image. However, LBP still has many disadvantages. For example, applied it in classification is not good enough, under illumination threat it is not robust.This paper considers the importance of the muti-resolution on the recognition. It found that the features of the image can not be found in one resolution, but can easily be found in another resolution. Therefore, I proposed FWTLBP algorithm that combines LBP with Fast Wavelet Transform. First, decompose the original image using FWT. We can obtain the low resolution image. And then calculate the result image using LBP operator. Though that we can get feature information in different resolutions and reduce computation by half. In addition. we can move the image noise.Meanwhile, this paper proposed a muti-gray level FWTLBP operator. With the gray level increased, the contour of the face in muti-gray level LBP image becomes more clearly that is quite beneficial for face recognition. Because from the perspective of biological vision, contour plays an important role in identifying, and after that is details.Chapter 4 of this paper analyze the experiment results using muti-gray level FWTLBP computed on ORL face database, compare the recognition ratio and the variance computed by the basic LBP operator and muti-gray level FWTLBP respectively, and found that the algorithm proposed in this paper can effectively improve the recognition ratio. To further verify the effect of this algorithm, it is applied on YALE face database, and the recognition ratio can be as high as 0.99. In these experiments, we can get the conclusion that for the face image position changed we define the decomposition series 0 and gray level larger can get good results, and for those expressive we define the decomposition series 1 and gray level 0 can get better effect.This experiment uses MATLAB as a development tool, write and debug multi-gray level FWTLBP algorithm codes. When the computer hardware configuration is in a good, choose MATLAB as a development tool for image processing will be more convenient.LBP can apply widely. It has been used in face recognition, texture analysis, moving object detection, metal surface detection and other aspects. Its powerful advantage will attract more scholars to further research LBP.