Research On Texture Spectrum Based Feature Space And Its Application For Image Recognition

In recent years, image recognition is one of the research hot spots in the fields of image processing, computer vision and pattern recognition because of its potential values in theory and application. Image recognition system is mainly composed by the following parts: preprocessing, feature extraction and feature classification. Feature extraction is the key of the system, on which the selection of the learning algorithm, learning efficiency and recognition rate are dependent to a great extent. This dissertation focuses on texture and face recognition.Texture Spectrum(TS) and Local Binary Pattern(LBP) are two powerful measures of extracting texture feature used to describe the local texture unit of image, showing excellent results. However, TS contains too many feature bins and leads to large storage and computational costs. As its simplified version, LBP neglects the vagueness and inaccuracy introduced by noise and the digital processes and extracts features with loss of some discriminant information. Corresponding solutions of these limitations are proposed in this dissertation. The primary work of this paper can be summarized as follows:Firstly, fuzzy local binary pattern is proposed. Fuzzy techniques are introduced into the original definition of LBP. The extended method is not sensitive to vagueness brought by noise and digital processes as the old one and more consistent with the vision characteristics of human eyes.Secondly, uniform texture spectrum andεLBP2 are constructed. They ameliorate the TS texture spectrum method from horizontal and vertical view respectively to reduce the feature bins and the complexity of storage and computation.Thirdly, Gabor texture spectrum is proposed as a new feature representation. Gabor wavelet transform is performed to the image at the beginning, then TS feature is extracted from the transformed image as the representation of the original image. Thus the transformed feature representation is more favorable for classification.Experiments are carried out to verify the extended algorithms above on standard databases, and the results show that the proposed methods are effective and feasible.