Research On Image Texture Feature Extraction Based On Local Pattern

In recent years,with the rapid development of artificial intelligence,the application of image texture features has become an indispensable part of the field of computer vision.The extraction and recognition of image texture features is the basic work of image classification and image recognition.With the continuous improvement of image acquisition technology and the widespread use of high-definition cameras and other tools,more and more image samples are collected in complex environments.People pay more attention to image feature extraction subject to various interference factors(such as light,angle,noise,occlusion,etc.).It is proved that compared with the global feature extraction of the texture feature,the local feature extraction can better represent the details of the texture image.For example,the Local Binary Pattern(LBP)algorithm and its related local description algorithms and image surface Shape Index(SI)have good ability to describe the local information of the image.This kind of algorithm is widely used in feature extraction.This paper mainly studies the LBP algorithm and its improved method,and discusses the image feature extraction method of texture image under the interference of different illumination,angle,attitude and other changing factors.Aiming at the defects of LBP algorithm and its improved method in the process of image feature extraction,a method based on image surface shape index is proposed.The specific work is as follows:Based on the analysis of the principle of traditional LBP,an enhanced local binary pattern algorithm based on feature fusion of local binary pattern and circular binary pattern RLBP(Ring Local Binary Pattern)is constructed.This algorithm can extract more abundant facial texture features,and the local facial texture features are fully extracted,which enhances the robustness of the algorithm.Through experiments such as segmentation,classification,noise and other experiments on common texture data sets,our improved algorithm has a higher recognition rate in many aspects than similar algorithms.Traditional LBP and its variants have unique advantages in image texture feature extraction and classification,but they still have some inherent defects: continuous rotation invariance and poor robustness to noise.In view of the limitation of continuous rotation invariance,a new texture descriptor SI-LCCP is proposed based on the surface shape index of LBP fusion image.The Shape Index(SI)is constructed by the principal curvature of the image surface,which has the characteristic of continuous rotation invariance.We combine this characteristic to improve our algorithm.At the same time,we combine our proposed descriptor with Local Concave and Convex Pattern,extract richer texture features to improve its robustness to noise,and add a variety of noises to test and compare.Experimental results show that it effectively improves the robustness of the algorithm to noise.It is verified on four commonly used texture image data sets,and the results show that our scheme has the advantages of strong viewpoint robustness,overcome rotation invariance,strong anti-noise ability and so on.The spectrum of Gabor filter has the Gaussian characteristic,while the natural image spectrum has the non-Gaussian characteristic.The spectrum of natural image distribution can not be completely obtained by using Gabor filter,and some non-Gaussian frequency information will be lost.Based on the study of this problem,we propose to use the Log-Gabor function to filter the spectral images with non-Gaussian characteristics,and use the Generalized Gaussian Model(GGM)to model the Log-Gabor transform coefficients in different scales and directions.We estimate the model parameters by using the Expectation Maximum(EM)method.Finally,experiments on recognition accuracy and cumulative matching characteristic curves are carried out in three commonly used texture image databases,and the results show that the recognition rate of Log-Gabor coefficient correlation algorithm is higher than that of other related algorithms.