Research On Wood Species Recognition Based On Deep Learning

China is a large timber consumer and also one of the countries with the largest total timber import and export trade.With the increasing frequency of global timber trade,in order to prevent timber smuggling,overfilling and illegal logging,it is necessary to accurately identify timber species.At present,artificial wood species identification methods have problems such as strong professionalism,heavy tasks,long cycles,high risks,and non-real-time characteristics,which cannot meet the requirements of real-time and high-efficiency timber species identification.This paper proposes a wood recognition algorithm based on improved residual neural network to recognize the types of wood cross-section macrostructure images.The main research contents and results include:(1)Research on recognition models of different species based on deep learning.A mobile phone with a macro lens was used to collect macroscopic structural images of wood cross-sections.On the same data set,six models of VggNet16,GoogleNet,DenseNet,ResNet50,ResNet101 and ResNet152 were trained,and 120 original wood crosssectional images were tested.,The average recognition accuracy rates were 70.9%,80.9%,82.8%,87.5%,91.7% and 90.1%,indicating that the ResNet101 model is more suitable for wood image feature extraction and species identification than the other five models.(2)Research on wood species recognition model based on improved ResNet101.Aiming at the problem of information loss in the input layer of the convolutional neural network and blurring of the edge of the wood cross-section macrostructure image,a block strategy and a gradient weighting algorithm are proposed to improve the ResNet101 wood species recognition model.The wood cross-section macrostructure image uses different gradient values from the edge to the center as the weight of the image classification scores of different sub-regions,increasing the proportion of the central region in the entire image recognition task,and the improved model features each sub-image Extract and calculate the final classification score of each image.(3)Training,testing and analysis of wood species recognition model based on deep learning.Using the same data set to train 5×5,7×7 and 10×10 three ResNet101 models based on different block strategies of the original image,the average recognition accuracy rates of 94.8%,96.1% and 95.3% were obtained respectively;the gradient was weighted The algorithm is applied to the ResNet101 model under the7×7 block strategy,and obtains an average recognition accuracy of 98.4% and an average recall rate of 98.7%,which is the same as the original image + ResNet101 and 7 × 7 block strategy + ResNet101.Compared with the recognition model,the average recognition rate increased by 6.7% and 2.3%,and the average recall rate increased by 7.4% and 2.8%,respectively,indicating that the block gradient weighting algorithm can effectively improve the accuracy of the wood recognition model.The algorithm proposed in this paper has good application and promotion value.