Research On Wood Tree Species Recognition Based On Lightweight Convolutional Neural Network

The effective utilization of wood resources and market transactions are based on accurate identification of wood tree species,but the traditional manual wood identification methods have problems such as long cycle,heavy tasks and many human disturbances.Therefore,exploring accurate and fast wood tree species identification methods is one of the important research contents in the field of wood.In recent years,deep learning technology has made remarkable progress,and it has been widely used in many fields,especially in the field of image recognition,and has achieved remarkable results.Based on the lightweight convolutional neural network algorithm in deep learning,this paper takes the three-section(cross section,radial section,and chord section)images of wood tree species as the research object,and constructs an improved lightweight wood tree species identification model,and built a graphic application based on this wood tree species identification model.The main research contents are as follows:(1)Build a dataset of images of wood tree species.The three-section image dataset of wood required for the research was obtained by two methods: on-site collection and network collection.The on-site collection was obtained by manual photography in the school’s wood specimen room;the network collection was obtained through the public dataset of Barmpoutis.Ten kinds of wood,including Turkish oak,Chinese fir,star fruit,cypress,walnut,European beech,European chestnut,French sycamore,rain tree,and boxwood,were selected as the dataset of this research,and finally the construction of the wood tree species image dataset was completed.The wood tree species image dataset was preprocessed by image size normalization,image grayscale and custom histogram equalization,and image data enhancement was achieved by image rotation mirroring and image random cropping to expand the wood tree species image dataset.(2)Build a wood tree species recognition model based on the improved MobileNet＿v3＿small network.The Research improved the activation function and the overall structure of the network in the inverted residual structure bneck of the lightweight MobileNet＿v3＿small network,and trained Res Net50,VGG16,MobileNet＿v2,MobileNet＿v3＿small,MobileNet＿v3＿large and the improved MobileNet＿v3＿small on the same wood tree species image dataset.The wood tree species recognition model was tested and compared using the image testset to test the performance of the improved network model.The test recognition accuracy rates are 99.23%,98.31%,98.87%,97.93%,99.04% and98.85%,respectively.Compared with the MobileNet＿v3＿small model,the improved model has a significantly improved recognition accuracy and a similar number of parameters.Compared with the MobileNet＿v3＿large model The accuracy rate is slightly reduced,but the amount of parameters and calculation is reduced by 50%.After comprehensively considering the accuracy rate,parameter amount and calculation amount,the improved model proposed in this paper has certain performance compared with other models in the wood species image dataset.The advantage is that it has the characteristics of high recognition accuracy and relatively small amount of parameters and calculation.(3)The GUI of the wood tree species identification system was constructed by using the Tkinter module that comes with Python,which includes a user login GUI interface and a wood tree species identification GUI interface.The GUI of the wood tree species identification system realized functions such as login and registration,image upload,wood tree species identification,and wood tree species introduction,providing a new way for forestry practitioners and related researchers engaged in timber research to identify wood tree species.