A Variant Of LBP For Texture Classification

With the rapid development of the network transmission and terminal equipment, imagehas become one of the main sources of information with an explosion growth. The vastamount of images challenges the information technology. How to effectively understand anduse the image data and how to extract useful information from huge image library havebecome one challenging and meaning research topic. Image classification, as a critical imageanalysis technique, has been widely used in intelligent video surveillance, target recognition,content-based image retrieval and other aspects. The texture characteristic is the main part ofimage features and plays an important role in image classification.Texture classification is a quite active research subject in computer vision and patternrecognition, etc. Local binary pattern (LBP), proposed in recent years, and is widely used fortexture analysis because of its robustness to illumination and rotation changes and lowcomputational complexity. But LBP feather is obtained locally with the defect of losing globalinformation. In addition, among the vast amount of images, the labeled images is difficultobtained, but the unlabelled is easy obtained. According to these problems, this paper putsforward a novel solution, Tri-LBP, based on Tri-training for the texture image classification.The main content includes:This approach combines LBP, LBPV and LBP-HF, three different features based onTri-training. It takes account of information both in local and global, so can fully describes thetexture characteristics. Meanwhile, the semi-supervised learning of Tri-training can improveclassification performance with little labelled sambles and large number of unlabelled ones.Our approach is evaluated by using texture images from the Outex database. Theexperiments show that the proposed method is effective for gray-scale and rotation invarianttexture classification. In addition, the classification accuracy can be further improved byincreasing the differences between the classifier.The comparison with other methods illustrates that our proposal obtains a betterperformance. The application of Tri-training improved the classification accuracy obviously.