基于改进DenseNet的福建常见阔叶材显微识别研究OA

Fujian Province in China is home to a wide variety of broad-leaved tree species and boasts abundant forest resources.To enable fast and accurate identification of hardwood species,this study proposed an improved DenseNet network model for the microscopic identification of wood species.Twenty-four common broad-leaved wood species from Fujian were selected as research subjects.Compared to macroscopic images,microscopic wood images provide more detailed and abundant information about wood structural characteristics.Microsections of the selected common broad-leaved wood species were prepared,and original microscopic images of their cross-section were collected.To reduce the computational complexity of image processing,the images were preprocessed using techniques such as image size normalization and image grayscale conversion.Additionally,data augmentation methods,including horizontal flipping,random scaling,image rotation,and adjustments to brightness,contrast,and saturation,were applied to enhance the diversity of the training dataset and mitigate overfitting.Through the above process,the microscope image data set of cross-sections of common broad-leaved wood species in Fujian was constructed.Four classical convolutional neural networks,i.e.,VGGNet19,InceptionV3,ResNet101,and DenseNet121,were trained on 24 kinds of microscopic images data sets for hardwood species in Fujian,respectively.The recognition accuracy,training time,parameter number and model file size of the four networks were compared and analyzed.It was found that the DenseNet121 model possessed the highest recognition accuracy(98.02％),the shortest training time(2.56×104 s),the least number of parameters(7.57×106)and the smallest model file(30 MB),indicating that the DenseNet121 model had the best overall performance among the four classical convolutional neural networks.The results showed that DenseNet121 possessed better overall recognition performance on this data set.The DenseNet121 model with the best overall performance was selected for improvement.The number of parameters in the network model was reduced by introducing deep separable convolution in the network except for the initial convolutional layer,and the recognition performance of the model was improved by introducing Inception module and channel attention mechanism.The result showed that the average recognition accuracy of the improved DenseNet model reached 98.96％and the average recall rate was 98.95％.The training time,parameter number and model size of the improved DenseNet model were reduced by 0.9×104 s,5.66×106 and 6 MB,respectively,compared with DenseNet121.The improved DenseNet significantly enhanced recognition performance while greatly reducing computational and storage requirements.The results offer a scientific and efficient method for wood identification personnel to accurately classify hardwood species.