Texture Image Feature Extraction And Classification

Texture feature extraction and classification are widely used in remote sensing,medicine,agriculture,industry,etc.,such as terrain detection,disaster prevention,crop surveillance,medical image analysis.Traditional texture feature extraction methods have some weaknesses:some texture features are sensitive to rotation,pose,viewpoint and scale variation,low classification efficiency,poor results in some applications.According to the problem of rotation texture feature extraction,the paper presents a new multi-scale rotation invariant texture extraction(MSRIT)method.The MSRIT is invariant to rotation,pose,viewpoint and scale variation.According to the low classification efficiency of traditional methods on texture images,the paper proposes a new SVM model:SVMpdip and a new method to solve the model:the primal-dual interior point method based on block elimination(PDIPbe).The SVMpdip can achieve higher classification accuracy and run faster than some traditional classification methods.The MSRIT and SVMpdip can classify complicated texture in practice.The MSRIT is extracted from the multiple rotation-invariant local feature descriptors at mult-scale images.It is invariant to rotation and scale variation.In terms of the SVMpdip,the block elimination is used to divide the sparse matrices produced during computation to some block matrices most of which are special,such as zero matrices,identity matrices,diagonal matrices.These special matrices are easy to store and compute,which save much memory,reduce computation complexity and improve classification efficiency.Furthermore,to accelerate the convergence a starting point is found for the PDIPbe.Based on the therotical analysis,the paper performs many experiments and evaluations on many well-known texture datasets.The experimental results of the MSRIT against the four well-known texture features:Gabor,GLCM,GLDM,LBP,the SVMpdip against six well-known classification methods:SMO-P,SMO-K1,SMO-K2,CVX,quadprog,svmlight on multiple datasets demonstrate that the MSRIT outperforms the other textures on classification accuracy,the SVMpdip mostly runs faster and achieves higher classification accuracy than the other classsifcation methods.