Study On Wood Species Recognition Based On Convolutional Neural Network

Wood species identification is a vital task in wood science and industry.Traditionally,the identification of wood species is done by people with professional knowledge and rich experience,who identify wood based on various characteristics of wood(such as color,structure,and texture).However,the manual recognition of wood species has low accuracy and is a time-consuming process.Therefore,it is necessary to develop a system for automatically identifying wood species.The traditional method of automatically recognizing wood species is to extract the features of the wood image with a manually designed extractor,and then use a machine learning classifier to identify the wood based on these features.However,this kind of manual feature extraction method is difficult to extract high-quality features,resulting in low recognition accuracy.To solve the above problems,the convolution neural network in deep learning is used to identify wood species,which can automatically learn the features of the wood section images and greatly improve the recognition accuracy.The dataset of wood species in this paper is composed of the images of the cross section,radial section and tangential section of 12 kinds of wood.Aiming at the problem that the number of samples in the wood species dataset in this paper is small,this paper performs data augmentation on the wood species dataset to prevent overfitting phenomenon of convolution neural network.In this paper,two wood species recognition methods are proposed based on the convolutional neural network.The existing classical convolutional neural network is used to identify wood species.To improve the generalization performance of the network,this paper uses the fine-tuning technology in transfer learning to train the classic convolutional neural network.In this paper,four classical convolutional neural networks are selected: Alex Net,VGGNet,Goog Le Net and Res Net.Through comparative experiments,it is proved that the finetuning technology of transfer learning can improve the recognition accuracy of these four networks in the task of wood species recognition.Among them,Res Net-50 trained by using fine-tuning technology of the transfer learning achieves the highest recognition accuracy of 99.15%.In this paper,a new lightweight convolutional neural network is proposed to achieve real-time wood species identification on a platform with limited storage space and computing power.This network is the result of a significant improvement in Mobile Net V2.The biggest improvement is the removal of 10 inverted residual blocks with shortcut connections,so the amount of calculation and parameters is greatly reduced.To improve the generalization performance,this network adopts the label smoothing strategy training and achieves a recognition accuracy rate of 99.22%.In this paper,this network,Mobile Net V2,Shuffle Net V2 and the above four classic convolutional neural networks are put together to compare the performance.The recognition accuracy,parameter amount and calculation amount of this network reach the optimal level.Finally,this paper designs and implements a wood species recognition system based on the Android operating system.The recognition speed and occupied space of the lightweight convolutional neural network proposed in this paper are superior to other networks in the paper.