Research On The Method Of Image Texture Feature Extraction And Classification

Image classification, object detection and recognition, and scene classificationdepend on the description and extraction of image characteristics. The texture is one of themost frequently used features by human visual system. Texture provides informationabout the spatial distribution and brightness transformation, and also describes thestructure arrangement on the surface of the image. However, despite the human visualsystem is easy to accurately identify and describe the texture, but it is hard to extracttexture features, used in texture classification as well as different application fields.At first, this paper studied the texture feature extraction algorithm, and then used theLIBSVM software developed by Lin, national Taiwan university for image classification.This paper made the following several aspects of research:(1) This paper studies the commonly algorithm of the texture analysis and classicalalgorithm of image classification, and summarized the basic ideas of texture imageclassification.(2) In view of the influence on classification of image rotation, this paper proposedthe rotational invariance of the Radon transform and wavelet transform algorithm. AfterRadon transform, the rotating image has translation relative to the original image. At thesame time, wavelet transform can make the translation not eliminate. Therefore,combining the Radon transform and wavelet transform, can realize rotation invariance oftexture image.(3) In order to achieve both multi-scale and statistical characteristics of image, graylevel co-occurrence matrix is put forward combined with the wavelet transform algorithm.In view of the way of combination, two different methods are put forward. One is thatfirstly process the image by wavelet transform, and then calculate the gray levelco-occurrence matrix of transformed images. And another is that respectively to dowavelet transform and gray level co-occurrence matrix, then extract respectively thecharacteristics, and combine the characteristics as texture image feature vector. The last isthe use of support vector machine (SVM) for classification.(4) The paper designed the process of the experiment to realize function module.Then do experiments about the effect of Penalty factor C and kernel function and itsparameters on the accuracy of classification. Finally compare all kinds of texture featureextraction algorithm, as well as the effect on the accuracy of classification under theinterference of noise. Through a series of the experiment the result shows that the improved algorithms have better classification effect compared with the classicalalgorithm, and have higher classification accuracy.