Research On Classification And Recognition Algorithm Based On Rosewood Microstructure

Rosewood,as a kind of precious timber,has high economic value,artistic value and research value.Rosewood is used as a connecting member or decorative material in ancient furniture and wooden structure building.As a part of modern furniture and decoration,rosewood is also used to improve environmental comfort.Accurate and rapid identification of rosewood species is of great significance and value for the restoration of ancient wooden structure building,wood species identification of rosewood furniture market and wood import and export trade.At present,there are chemical methods,DNA methods,spectral identification methods and anatomical methods,which are used for the identification of rosewood species.The study of rosewood identification based on image mainly uses anatomical methods.There are two main problems in this method: First,it is difficult to obtain microscopic images of rosewood.Wood was made into slices after sampling,dyeing and dehydration,and then observed by microscope.Second,rosewood species are identified more by experienced experts than by computer visionbased methods.The research in this thesis has solved the automatic identification problem in rosewood identification,from the aspects of the acquisition of rosewood microscopic images,multi-dimensional texture feature fusion,convolution neural network,and vessel pores feature extraction.The research contents and innovations of this thesis are as follows:（1）The microscopic images of cross sections,radial sections and tangential sections of rosewood were reconstructed by micro-CT,which created a digital specimen library of rosewood for wooden structure building materials museum of Shandong Jianzhu University,and laid a foundation for the study of rosewood species identification based on computer vision method.（2）In view of the limitations of the identification method of single texture feature of rosewood,according to the characteristics and laws of different microstructure distribution of rosewood,this thesis proposes a multidimensional texture feature fusion method for rosewood identification.Five feature fusion methods were constructed by fusing single LBP（Local Binary Patterns）features and deformation forms with GLCM（Gray-level Co-occurrence Matrix）and Tamura features,respectively.The classification and recognition of the cross sections,radial sections and tangential sections images of rosewood were realized by combining radial basis function,BP neural network and extreme learning machine,respectively.The experimental results show that the classification accuracy of feature fusion is generally higher than that of single feature.（3）In order to explore the influence of various section combinations on the accuracy of rosewood identification,a convolutional neural network model（Con Net Model）is constructed,and the cross sections,radial sections or tangential sections images of rosewood are used as model input to realize the identification of wood species.The inception module uses multiple convolution kernels of different sizes to obtain multi-scale features,and adds a scaling layer to the model before the Leaky Re LU activation function and on the jump line of the residual block.Compared with the traditional classification models,the identification accuracy and classification efficiency of this model is significantly improved.One section,two section combinations and three section combinations of rosewood specimen were used as the input of Con Net Model to train the rosewood identification model of convolutional neural network and compare the effects of different types of sections on the accuracy of wood species identification.After comparing many experimental results,the accuracy of single section identification is higher than that of two or three section combinations.（4）By observing the characteristics of different structural distribution on the cross-section microscopic images of rosewood,a recognition method based on vessel pores characteristics is proposed.Firstly,a vessel pores segmentation model（Onet）is constructed,which consists of two ’ U ’-shaped network structures.Two ’ U ’-shaped structures accept the same input.Through feature extraction and feature fusion with different convolution operations,the accurate segmentation of vessel pores is realized.Secondly,the Dual Resnet18 was proposed for the identification of vessel pores characteristics.The model introduces the upper and lower parallel residual block structures,and introduces the gain processing before the Re LU activation function and on the residual block jump line.Compared with the classical model,this model has significant advantages in rosewood species identification.In summary,the multi-dimensional texture feature proposed in this thesis integrates the rosewood identification algorithm,the convolutional neural network rosewood identification algorithm and the vessel pores feature rosewood identification algorithm,which realizes the intelligent classification and identification of rosewood and the systematization of "machine expert" decision-making.In the identification process of wooden structure building component species,the accuracy needs to be guaranteed.In the process of repairing building components,the principle of equal proportion replacement of the same material species needs to be respected.The original architectural style and characteristics of the times need to be preserved.The buildings and cultural relics need to be restored as old.