Study On Feature Extraction Method Of Texture Image Based On Principal Curvatures

With the rapid development of the digital multimedia information era,large numbers of images and videos are continuously spread and shared on the computer network.The high-dimensional nature of these videos and images information often leads to time complexity and space complexity in the imaging process,resulting in a "dimension disaster".Along with the "dimension disaster" problem of multimedia data,a large amount of irrelevant or redundant information exists in high-dimensional space,which seriously affects the content analysis and processing of images and videos.Therefore,timely acquisition and effective processing of information in videos and images are research hotspots in the multimedia era.As an important visual cues of the image surface,the texture features reflect the surface properties of the scene corresponding to the image or image area,and play a vital role in the field of image processing,pattern recognition and computer vision.However,natural texture images often produce large differences in imaging results due to changes in external imaging conditions.Therefore,to process the information of image in time and efficiently,it is of great theoretical significance and application value to study the simple,efficient,low-dimensional and robust image texture feature extraction method.The paper mainly focuses on the texture feature extraction method of image.For the existing problems in the image texture feature extraction process,based on the principal curvatures information at the pixel points of the image surface,four simple,efficient and robust texture feature extraction methods are proposed by using feature fusion,multiscale and multiresolution analysis,cross-scale joint feature description.The main research contents and innovations of this paper are as follows:(1)To solve the influence from rotation and illumination changes during texture image imaging,an image texture feature extraction algorithm based on complete local binary pattern and principal curvature is proposed.Image texture features in real-world environments often suffer from arbitrary rotation and changes in illumination,which can severely impact the results of statistical methods.The principal curvatures information at the pixel of image surface not only has continuous rotation invariant characteristics,but also contains two extreme curvatures that can effectively describe the macro and micro features of image,which is very effective for solving the influence from to illumination and rotation changes.Based on the above advantages of the principal curvatures information,this paper transforms and encodesthe principal curvatures and fuses them with the feature information obtained by complete local binary pattern.The experimental results demonstrate that using the principal curvature as the fusion feature information can effectively improve the robustness of the complete local binary mode to illumination and rotation changes.At the same time,it lays a foundation for the more robust and efficient image texture feature extraction algorithm proposed in the next chapter.(2)To solve the poor robustness to rotation,high dimensionlity and low accuracy of the existing algorithm during the processing of image texture feature extraction,This paper proposes a multiresolution gray-scale and rotation invariant feature extraction algorithm based on principal curvatures.Although many feature extraction algorithms have been proposed and can achieve certain effects,their computational complexity and feature vector dimensions are very high.Moreover,the rotation invariant patterns constructed by these methods are not completely rotation invariant.To solve the above problems,this paper designs a more efficient feature information conversion and normalized method based on the principal curvatures of the pixel points on the image surface,and encodes the converted and normalized results.At the same time,only the symbol information of the local pixel is used as the fusion information fused with the encoded principal curvatures information.Moreover,to obtain the multiresolution texture feature information which is simple,efficient and robust to illumination,rotation and viewpoint changes,we design a multiresolution information extraction scheme with different parameters and superimposed the extracted multiresolution feature information.Then,through experimental simulation and parameter optimization,the algorithm is optimized to determine the detailed parameters with optimal performance.The experimental results demonstrate that the proposed feature extraction algorithm outperforms the existing image texture feature extraction methods in terms of feature dimension,accuracy and robustness to illumination,rotation and viewpoint changes.(3)Image texture is susceptible to scale scaling and viewpoint change,and the information extracted by the existing texture feature extraction algorithm is redundant and low-discriminating.Furthermore,their performance is limited by the number of training samples while classifying and identifying texture targets.In this paper,a cross-complementary local binary pattern feature extraction algorithm based on principal curvature is proposed.On the basis of extracting,normalizing,transforming and encoding the principal curvatures of the pixel points on the image surface,wepropose a multiscale and multiresolution feature information extraction scheme.Moreover,to achieve robust extraction of image texture information,the cross-scale joint feature description method is used to fuse the principal curvature coding information and the rotation-invariant unified pattern local binary coding information at different scales and resolutions.The experimental results demonstrate that the proposed algorithm is not only robust to scale,viewpoint,illumination,rotation and sample number changes,but also has simple,high-efficient and low-dimensional characteristics,which can effectively extract the feature information of the texture image.(4)Gaussian curvature and mean curvature in the curvature information at the pixel points on the image surface are also researched in this paper.According to their relationship with the principal curvatures,a new image texture feature extraction algorithm based on Gaussian curvature and mean curvature is proposed.In the proposed algorithm,the Gaussian curvature and the average curvature are calculated by Gaussian first-order derivative and second-order derivative,and then transformed into the principal curvatures information and subjected to feature transformation and normalization.Subsequently,the normalized result is encoded in the local binary pattern.Furthermore,to extract the more robust feature information,multiscale and multiresolution scheme as well as cross-scale joint feature description method are used to combine the rotation-invariant unified pattern local binary coding information with the principal curvatures coding information.Finally,through simulation and optimization,the detailed values of each parameter when the algorithm achieves optimal performance are determined.Compared with the current mainstream texture feature extraction algorithm,the experimental results demonstrate that the proposed algorithm is highly robust to the changes of various imaging conditions,and can extract the texture feature information effectively,which provides another simple,low-dimensional,and high-efficient method for the feature extraction of texture image.