Research On High-Performance Texture Extraction Method Based On Local Binary Pattern

Texture is a key visual cue for humans to perceive different objects,and texture recognition plays a crucial part in the area of computer vision.Among plenty of texture feature extraction methods,Local Binary Pattern(LBP)has received extensive attention from researchers because of its simplicity and effectiveness,invariance to gray scale,and no training and learning.Although many LBP-based texture extraction methods have obtained good classification performance,the current LBP algorithms are still sensitive to the illumination and rotation of the image,and are not robust enough to image noise,resulting in an unsatisfactory discrimination ability.To deal with the above problems,the following research work is conducted in this article:(1)Aiming at the fact that the Local Grouped Order Pattern(LGOP)ignores the difference information between neighborhoods,and the robustness to external Gaussian noise needs to be improved,an enhanced local grouped order pattern and non-local binary pattern is proposed.First,a first-order difference coding scheme is proposed,which encodes the neighborhood difference information of groups based on symbol difference to obtain a more complete representation of neighborhood information.Since the gradient images contain rich structural information,this paper also performs cross-image domain information fusion,that is,fusing the texture information of the original image domain and the gradient domain.This paper conducts extensive experiments and evaluates noise robustness on four mainstream datasets(Outex,CURe T,KTH-TIPS,UMD),and the experimental results prove that the proposed texture descriptor has better classification performance than existing LBP algorithms with and without noise.(2)To further enhance the robustness to noise and the performance of texture classification,this paper designs a weight function to give greater weight to the grouping of dissimilar neighborhood sampling points,and proposes a weight-based local grouped order pattern for texture description.This method improves the robustness of the LGOP descriptor to noise by computing the variance of each group of neighboring sampling points,finding the sampling points whose variance is greater than the average variance,and assigning greater weight to the local grouped order pattern code value corresponding to these sampling points.Through the analysis of the classification experiment results of Outex(TC10,TC12),CURe T,and KTH-TIPS texture databases,it is found that the method can achieve good texture classification performance in both noise-free and Gaussian noise environments.(3)To improve the discrimination ability of texture features and the ability to resist image noise corruption,this paper calculates two complementary features,the gray level co-occurrence matrix statistics of different image domains and the joint shape index of different scales,and then combines them with Local Contrast Pattern(LCP),a local contrast pattern combining multiple features is put forward.First,the gradient image is generated by convolving the original image with the Sobel operator,and then the gray level co-occurrence matrix statistics of the two images are calculated.At the same time,the shape index is introduced and quantized to obtain discrete texture codes.Finally,the above three types of features are jointly encoded with histogram to form the final texture feature expression.The classification experimental results on the datasets Outex TC10 and CURe T verify the good noise robustness and discriminative ability of our proposed method.(4)Based on the above work,a system prototype based on Qt framework is realized.