Research And Application Of Improved LBP/LTP Algorithm

As a kind of images texture features descriptor,the local binary pattern can extract the texture features of images simply,effectively and quickly.In order to improve the anti-noise ability and robustness of the LBP algorithm,scholars have proposed many improved algorithms.As an improved algorithm of LBP,local ternary pattern(LTP)introduces a threshold to encode images,which effectively improves the robustness of the binary algorithm to illumination variation and noise interference.The noise-resistant local binary pattern(NRLBP)corrects the non-uniform patterns in the LBP algorithm by introducing ambiguity.However,the above two improved LBP algorithms' thresholds are fixed values and lack of adaptability.In this paper,the research on improving the classification and recognition accuracy and anti-noise ability of the existing LBP algorithm is carried out.The main research contents are as follows:(1)The threshold of LTP is fixed,which greatly increases the sensitivity to noise,and cannot achieve the adaptability of texture feature extraction.A multi-scale adaptive threshold local ternary pattern algorithm is proposed.First,multi-scale transformation of the image,divide images of each scale into several regions,and the average value of pixels in each region is calculated;Then extract the standard deviation of each pixel from the mean of the extraction region and use it as the adaptive threshold;Finally,the difference between the neighborhood pixels and the regional mean is compared with the standard deviation,the three-valued feature extraction histogram is obtained,and train the support vector machine to classify the texture images.Through the classification experiments of KTH-TIPS,CUReT,Outex,UIUC,Brodatz standard images libraries and railway fastener images libraries,the recognition rates of the improved algorithm is 0.59%-7.50 higher than the original LTP algorithm under different noise.Theoretical analyses and experimental verification show that the MSALTP algorithm selects different thresholds according to the intensity of each region pixel,effectively solving the noise sensitivity problem;The MSALTP algorithm solves the difficult problem of the fixed threshold determination of the original LTP algorithm;Multi-scale analyses of images can better describe local texture information and has higher recognition rates.MSALTP algorithm has better robustness and adaptability.(2)The threshold of NRLBP is fixed,which increases the sensitivity to noise and cannot achieve the adaptive problem of texture feature extraction;And using NRLBP to forcibly change the number of binary transformations more than two times into uniformpatterns will result in the loss of certain features;A multi-scale adaptive threshold enhanced noise-resistant local binary pattern is proposed.First,images are multi-scale transformed,and the image of each scale is divided into several regions,and the pixel mean value of each region is calculated;Then extract the standard deviation of each pixel from the mean of the extraction region and use it as the adaptive threshold;Comparing the difference between the neighborhood pixels and the regional mean with the standard deviation,and deriving the uncertain code;Then determining an indeterminate bit based on the value of the adjacent positions,and selecting a pattern in which the number of changes of 0/1 or(1/0)is not greater than four;Finally,the histogram training support vector machine is extracted to perform texture images classification.Through the classification experiment of the standard images libraries and the railway fastener images library,the recognition rates of the MSAENRLBP algorithm is increased by about 2.17%-13.10% under different noises.Theoretical analyses and experimental results show that the MSAENRLBP algorithm proposed in this paper can not only effectively solve the problem of important texture information loss such as lines,but also solve the problem that the NRLBP algorithm determines the threshold value.Multi-scale analyses of images can better describe local texture information.MSAENRLBP algorithm can better describe images and more local texture information with higher recognition rates;it has better robustness and adaptability against noise interference.