Application Research Of Rosewood Identification System Based On Convolutional Neural Network

Rosewood furniture has high market prices due to its scarcity of raw materials and high collection value.In order to make high profits,some unscrupulous businessmen use fake and shoddy wood instead of rosewood.At present,in addition to the traditional identification of rosewood by relying on the professional knowledge of professionals,the main identification methods are based on NIR spectral detection,DNA detection based on gene sequencing,proton NMR spectral detection and THz-TDS spectral detection,etc.Compared with the traditional manual inspection method,the existing inspection methods solve the problem of the destructive inspection of rosewood to a certain extent,and realize the micro-destructive inspection of rosewood.However,the existing detection methods still have problems such as low detection accuracy,micro-destructive detection and inconsistent detection standards,and these detection methods are limited to researchers.Aiming at the current problems of rosewood identification,this paper proposes a new method for wood identification based on convolutional neural network,and develops a software system.This method uses X-ray 3D CT to obtain the internal organization and structure of wood,and learns the characteristics of wood by training a 3D convolutional neural network,so as to realize CT image recognition of wood.For the convenience of users,this paper integrates the algorithm into the open source software framework Voreen and designs a visual interface for human-computer interaction.This article main research content is as follows:(1)Obtain high-quality 3D images of wood by X-ray 3D CT,solve the problem that the recognition accuracy of rosewood is limited by the accuracy of the detection method,and at the same time realizes the non-destructive detection of rosewood and establishes a unified detection standard.Through testing with different scanning parameters,it is finally determined that the CT scan can obtain a clear internal organization and structure of the wood under the conditions of voltage 60 k V,current 60 m A,spatial resolution 1.5um,exposure time 0.5s and scanning frame number 1440.It laid the foundation for the subsequent detection of image recognition.(2)The CT image of wood is classified by 3D convolutional neural network,which provides a new solution for wood CT image recognition.First,a 3D convolutional neural network suitable for wood CT image recognition is built by combining the advantages of residual ResNet and the conventional 3D convolutional neural network.Then increase the data through tailoring and data augmentation,thereby enhancing the generalization capability of the network model.Finally,the classification results of different algorithms are compared to verify the validity 3D convolutional neural network proposed in this paper.At the same time,the network hyper-parameters were optimized by controlling the variable method.(3)Through the open source software framework Voreen,the deployment of the convolutional neural network model and the design of visual interface are achieved,which is convenient for users to use the algorithm in this paper for wood detection.First,the algorithm model is integrated into the open source software framework by calling the C++interface of tensorflow to realize the forward inference of the network model.Then build a data flow network on Voreen to realize the visual interface of human-computer interaction.