The Resarch And Application Of Texture Feature Based On LBP

Texture feature is one of the most important image features,our paper mainly studies the Local Binary Patterns algorithm which evolves from the texture spectrum method for representing texture feature of image.Our improved algorithm is applied to image classification and recognition,target tracking and image segmentation.The main contributions of this paper are as follows:1.Through the study of LBP pattern classification method of uniform pattern and rotation invariant uniform pattern,we proposed a new pattern classification method that the pattern was classified by counting the number of value 1 's in the binary neighbor sets and the times of 0/1 transform.The proposed pattern classification method shows high texture recognition ability through histogram of image and texture library.2.Some methods of image processing were analyzed and classified by symbiosis theory in biology.For the texture feature extraction method of Pairwise Rotation Invariant Co-Occurrence Local Binary Pattern,to overcome the short of high computational complexity,poor rotation invariance and insensitive in small texture structure,an improved Pairwise Rotation Invariant Co-Occurrence Local Binary Pattern is proposed.First,the change of the selection of Co-Occurrence points in the improved algorithm enhances the robustness of rotation invariance and illumination change while maintaining the ability of the original algorithm to calculate higher order texture information.Secondly,the improved algorithm also increases recognition capability of images with small texture because of combining local pixel texture information.Moreover,the improved algorithm has less computational complexity than original algorithm.Through the experiment of Brodatz,Outex,CUReT and KTH TIPS image library,the recognition ability of the improved method compared with the original algorithm has rose by 0.17%,0.24%,2.39%and 2.04%.The experimental results show that the improved algorithm has better recognition for the image with small texture structure.The LBP is sensitive to noise and rotation variance.In order to solve this problem,a texture classification algorithm based on noise-tolerant co-occurrence local binary pattern was proposed.At first,the patterns of LBP are reclassified to extend uniform pattern and rotation invariant uniform pattern.Then,the local binary pattern which stands for visual micro-structures of original image and local binary count pattern which represents non-visual micro-structures of down sampling image are connected in parallel and image gradient magnitude information is added to form a new method that is robust to rotation variance and noise.In the end,the recognition rate of our method and other algorithm is compared in different texture image library.The experimental results show that our method has good rotation invariance and noise resistance.3?To solve the problem that Spatio-Temporal Context Learning algorithm can not trace the target in the case of occlusion and scale change,we put forward a new Spatio-Temporal Context Learning algorithm which fused LBP texture feature.Firstly,calculating the LBP texture histogram for each frame of the target region.Secondly,compute the similarity of target area LBP texture histogram between the first frame and the current frame by chi square statistics.The similarity of target area LBP texture histogram of two adjacent frames was also calculated.When the similarity is greater than the threshold value,we considered the target was occluded and we should change the updating coefficient of the temporal and spatial context and the central position of the tracking target.The experimental results show that in the requirement of real-time,the proposed algorithm is more robust to the target tracking in the case of occlusion and scale change compared with the STC algorithm.4?Aiming at the problem that the Region-Scalable Fitting Model is sensitive to the initial contour and has low segmentation efficiency,we proposed a new active contour model which mixed with LBP textural features.By introducing the texture energy term,the Region-Scalable Fitting Model is robust to the initial contour and has fast segmentation speed.