Research And Application Of Local Texture Extraction And Recognition Algorithm

As an important feature extracted from image data,local texture can provide key physical properties for further analysis,such as the differences in surface reflectance and so on.The repeated occur of local texture composes the global texture mode of image.Local texture feature is widely used in kinds of fields such as face recognition,license plate recognition,and industrial products detection.With the development of computer image and image analysis,texture features play more and more important roles.And the requirements of the texture-based application are higher and higher.Thus the higher requirement is called for the local texture extraction.For example,the texture-based description operator can adapt to various environment changes and have anti-noise performance and invariance to illumination,rotation,scale and view changes.Furthermore,accurately express the essence of image is required to improve the accuracy of the image recognition and classification.Therefore,the research of this paper include s the following aspects.First,the anti-noise performance of local texture feature is studied for the defects of local binary texture features.A local texture feature coding algorithm for adaptive noise suppression is proposed.Next,this paper studies how to express the local texture feature completely and propose a new local binary coding algorithm.A binary feature point description operator is proposed for the local feature point expression.Finally,combined with the engineering application,this paper studies the railway fastener detection approach.Local binary pattern texture algorithm is widely applied in image recognition.But it is sensitive to noise.A local texture feature coding algorithm for adaptive noise suppression is proposed for this defect.The image intensity is the result of measuring the illumination intensity of an object.In the measuring sense,noise is the gross error of the measurement results deviating for the real results.The noise intensities in each neighborhood are different.Thus,to insure no information is lost after filtering,the corresponding filtering parameters for each neighborhood are adopted.In any neighborhood,the image intensity is pretty similar or related.First,the proposed algorithm calculates the error between the pixel value and the mean value.Next,combined the theory that the random error is subject to Gaussian distribution,a criterion for automatic determination of noise point in the neighborhood is proposed.Then the noises are replaced by the mean values and the adaptive for local image is realized.Finally,algorithm tests are carried in the open image database and the testing results show the proposed algorithm owns good noise suppression performance.The original LBP or variants cannot provide enough sampled points,which cannot express the complete joint difference relation in the neighborhood.Gaus sian-based RSLBP(Random Sampling Local Binary Pattern)is proposed to solve this problem.The original and other improved LBP operator only expresses the joint deference relation between the central pixel and the pixels on the circumference of a neighborhoo d.The sampled points are not enough and the sampled positions are fixed.Thus the expressed joint difference relation is not complete.According to the pixel intensity follows Gaussian distribution,first,the proposed algorithm uses Gaussian distribution to determine the sampling points in the neighborhood.Then,it randomly chooses the point pairs and calculate difference between the point pairs.Finally,the joint difference distribution in the neighborhood is given,and the difference values of the poin t pairs are used for coding to get the local binary encode.The proposed algorithm increases the sampled points.So the local texture information is much richer.In addition,the original pixel distribution relationship in the neighborhood is considered and the sampled points can represent the original image neighborhood more effectively.The proposed algorithm encodes the local binary patterns by the difference between random pairs of points.The encoding result can better reflect the microscopic texture st ructure of the image.Experiments are carried on open image library and on-site captured fastener images.Theoretical analysis and experiments show that the proposed algorithm has better classification performance.The difference size and difference amplitude of pixel intensities in a neighborhood can completely express the difference relationship.For the feature point matching,a feature point description operator that fuses the difference size and the difference amplitude magnitude is proposed.The proposed algorithm first obtains sampling points according to the Gaussian distribution in the neighborhood;then randomly selects point pairs,calculates the difference size and differential magnitude of the pixel intensities between point pairs,and uses the binary model to represent the relationship between differences and the magnitude of the difference amplitude;finally,the relationship between the difference size and the magnitude of the difference amplitude of the pixel intensities is merged,and the description operator of the feature point is obtained.The feature point description operator obtained by using difference size and difference amplitude has higher discriminability.In addition,for the case where the difference in gray value is small and susceptible to noise,the pair of points within the threshold range for the difference is recalculated using its neighbor average,by setting the threshold.And the resulting code has better noise suppression capability.Aiming at the problem of rotation invariance of operators,thip paper proposes to calculate the gradient direction of pixels in the neighborhood,and use the gradient amplitude as the weight to obtain the gradient histogram,determine the maximum gradient direction as the main direction,and then rotate the neighborhood pixels to the main direction.Then,the difference calculation is performed to obtain the feat ure point description operator and the rotation invariance of the operators is achieved.Experiments were conducted on the open image library,which shows that the proposed algorithm has better noise suppression performance,rotation invariance,and better matching accuracy.For the low detection accuracy of railway fasteners images collected on site,RSLBP coding and LDA(Latent Dirichlet Allocation)are combined to detect the fasteners.Fasteners images are collected on-site in real-time.The noise and illumination impact the images greatly,which cause difficulty for fasteners detection.The proposed approach first process RSLAP coding to the images,then use LDA to extract the theme of the coded images and get the theme model of the image.Finally,the SVM was used to classify the image.The LDA theme model has a good performance to express the image theme,RSLBP can well reflect the local microstructure of the image,and the encoded image has a higher degree of discrimination.The combining the above two algorithms can reach higher detection accuracy on fasteners.Compared with other algorithms,testing carried on real fasteners images shows that the proposed method has higher detection accuracy.The paper concludes in the final conclusion and points out the further research content.