Research On Texture Feature Extraction And Automatic Classification Algorithms

As an important topic in the computer vision field, texture classification has brought great influence in numerous fields, and whether we can extract better features to distinguish different textures is the key factor. Different from early methods, as a new multi-resolution analysis tool, wavelet is appropriate for texture image processing for its perfect spatial-frequency localization performance and diversity. As a new developed texture analysis method, local binary pattern(LBP) can effectively overcome the rotation and non-uniform illumination problems for its rotation and gray-scale invariant character. More over, as a new machine learning technique based on statistical learning theory, support vector machine(SVM) is widely used in recent years for its perfect performance in processing classification problems in high dimension space.This thesis analyzed several better texture feature extraction algorithms. Besides that, some research were performed for texture image classification with SVM. The main contents and contributions are as follows:(1)A texture classification approach based on the dual-tree complex wavelet transform (DT-CWT) and support vector machine (SVM) was proposed. In this method, we utilized the DT-CWT with approximate translation invariance and good direction selectivity. Firstly, the texture images were decomposed into multilayer sub-bands via DT-CWT. Then the feature vectors were computed with mean and standard deviation of each sub-band image. Finally, the support vector machines were used for the texture classification. The experiment results showed that better accuracy for texture classification could be gotten with the proposed method.(2)As the rotation invariant texture image is concerned, a classification approach based on advanced LBP and SVM was realized. Commonly, the log-polar transform was used to convert the rotation to translation, and then the influence of rotation and scale change was reduced. Different from log-polar transform, we directly extracted gray-scale and rotation invariant features from the original images with the advanced LBP. The experimental results showed that this method could effectively solve the problem of gray-scale and rotation invariant.(3)We applied texture analysis method to the morphology detection, classification and identification of low-dimensional nanomaterials in this thesis. Based on the SEM images of nanomaterials, the texture features of materials were extracted by wavelet packet decomposition. Besides that, the nanomaterials were classified by combining the texture features with SVM. The experiments with 16 kinds of nanomaterials indicated that the classification accuracy rate tended to be 93.75%, so the validity of this method in practical engineering is verified.