Abstract:

Objective: A microscopic image recognition model TimberIDNet75 is proposed in this paper for 75 types of South American imported broadleaf timber, which provides an accurate method of multi-species recognition for customs, import and export quarantine inspection, and personnel engaged in timber identification research. Method: The TimberIDNet75 model is a medium-shallow neural network with 34 convolutional layers containing one input layer and four hidden layers. To expand the receptive field as much as possible to extract more image features, the input layer uses a 13×13×256 convolutional kernel to extract 256 types of features for each image, which are activated and pooled as the output. In the first hidden layer, two convolutions, activation and then residual correction are used, which is called“two-convolutions-one-correction-block”(TCOCB). The first hidden layer contains 3 TCOCBs, and 256 types of features are extracted as output. Then the second hidden layer contains 4 TCOCBs, extracting 512 classes of features as output. The third hidden layer contains 6 TCOCBs extracting features from the previous layer’s output, obtaining 1 024 classes of features. The fourth hidden layer contains 3 TCOCBs, after feature extraction from the output of layer 3, the features of 2 048 species were obtained, which were input into the fully connected layer after global average pooling to map the classification of 75 species. Result: The TimberIDNet75 model has an accuracy of 99.4% with a loss value of 0.044. Comparing the TimberIDNet75 model with the existing advanced deep learning models, ResNet model has an accuracy of 98.1%, VGGNet model has an accuracy of 97.1%, GoogleNet model has an accuracy of 96.2%, AlexNet model accuracy is 94.7%, ViT model accuracy is 53.2%, and TimberIDNet75 model accuracy is improved by 1.3% compared to the ResNet model, which has the highest accuracy among them. Then the TimberIDNet75 model was used to carry out practical tests on 75 randomly obtained microanatomical samples of imported broadleaf timber, and all the samples were accurately identified, with an accuracy rate of 100%. Conclusion: The TCOCB in the TimberIDNet75 model can eliminate the overfitting problem caused by model gradient drop while saving machine resources, meanwhile, the residual method makes it possible to minimize the manual intervention during model training, and the accuracy and efficiency are greatly improved.

Key words: wood microanatomy, deep learning network model, wood species identification, wood microanatomy feature extraction