Orientation Of Patch Based Rotation Invariant Texture Classification

Texture analysis is a basis issue in the field of image processing.With the development of texture analysis in the actual application,invariant texture analysis plays an important role in many fields.Rotation invariant texture classification means that the textures from different angles of the same category should be classified as the same category.It is an important direction of texture analysis,which has been widely used in many fields such as military,medical,object recognition,remote sensing and image retrieval.The study of rotational invariant texture classification will greatly promotes the development of texture analysis in various fields.Although the study of rotation invariant texture classification has been more than 30 years,the shortage of poor robustness,low classification accuracy,computation-intensive and time-consuming still exists.Among many algorithms,the patch based rotation invariant texture classification method missing direction information.This paper investigates the application of orientation information in patch based rotation invariant texture classification algorithm.The main work is follows:1.Rotation invariant texture classification using local rotation.In the feature extraction stage,we extract the feature pixel from texture patch by two ways,and use the maximum value of texture patch as the local orientation.To get the rotation invariant property and avoid the dimension disaster problem,we combine the rotation processing with random projection,which rotate the patch and project the high-dimensional feature into low-dimensional one.In the classification stage,using the concept of texture textons,we build the textons library from the low-dimensional feature by clustering,and place the textons distribution histogram in classifier for training and classifier.The simulation experiments are performed on four benchmark texture databases,UIUC,KTPTIPS,BRODATZ and OUTEX datasets.The results demonstrate the proposed method is better than some the-state-of-art techniques.2.Rotation invariant texture classification using principal direction estimation.In this paper,to keep the relationship between texture patches,the principal component analysis is applied to its local patch to estimate the local orientation,and the dominant orientation is determined by the distribution of local orientation.Then each local patch is rotated along the dominant orientation after circular interpolation.Taking into account the characteristics of the texture patch problem of high dimensionality,by using the random projection,the local gray value vector of a patch is mapped into a low dimensional feature vector together with local orientation feature.And the sample textons distribution histograms are placed into the classifier for training and classifying.The simulation experiments on standard texture databases OUTEX and CURET demonstrate the proposed method has a comparable performance with the existing methods and higher classification stability.