Abstract:

Objective: The traditional manual identification methods of wood species have some problems, such as strong professionalism, heavy task, long cycle and non-real-time performance and so on. A wood identification algorithm based on improved residual neural network was proposed to meet the requirements for real-time and efficiency of wood supervision. Method: Polished cross section images from thirty-two species of wood were processed. Firstly, a total of 8 975 cross-section images of wood were collected by mobile phone with external micro-lens. The gain coefficients were calculated by the means of average gray values of R, G and B channels. The product of each component gray value and the corresponding gain coefficient was used to replace the original gray values of three channels for eliminating the influences of color deviation caused by different image acquisition equipments and environments. Secondly, based on the self-similarity of wood cross-sectional macro structure, more training samples and image features were obtained by horizontal flipping, vertical flipping, addition of impulse noise and image block to ensure the relative balance quantity of different wood images. Then, the bilinear interpolation method was used to scale each sub-image to the same image size of 224×224 pixels. The improved residual convolutional neural network ResNet101 model based on block gradient weighting was used to extract the features of each sub-image, and the final identification score of each image was calculated. Finally, the average accuracy and recall rate were used to evaluate the identification results of different block processing strategies, different models and the improved residual convolutional neural network model. Result: On the same test set, VggNet16, GoogleNet, DenseNet, MobileNetv3, ResNet50, ResNet101 and ResNet152 models were used to identify the original cross-sectional images of 32 similar wood species, and the average identification accuracy rates were 71.3%, 81.3%, 83.2%, 66.4%, 87.9%, 92.1% and 90.5%, respectively. The ResNet101 model was selected to extract wood image features and identify wood species. Based on the ResNet101 model with 5×5, 7×7 and 10×10 blocks in the original images, the average identification accuracies of 94.8%, 96.5% and 95.3% were obtained respectively. The block gradient weighting strategy was applied to the ResNet101 model, and the average identification accuracy of 98.8% and the average recall rate of 99.1% were obtained. Compared with the ResNet101 models based on the original images and 7×7 blocks, the average identification rate of the ResNet101 model improved by using the block gradient weighting method was increased by 6.7% and 2.3%, and the average recall rate was increased by 7.4% and 2.8%, respectively. The block gradient weighting method can effectively improve the identification accuracy of woods. Conclusion: The 32 similar wood species were identified based on the ResNet101 model with block gradient weighting method, and the average identification accuracy was 98.8%. Wood cross-sectional images can be used to identify wood species, and the block gradient weighting strategy can improve the model identification rate.

Key words: wood identification, wood cross-sectional image, residual convolutional neural network, block, gradient weighting