Abstract:

The rise of deep learning and transfer learning provides a new direction for tree species identification. However, there are still challenges in wood identification among different qualities within the same tree species. In order to improve the quality grading status of zither panel, a deep residual network model is designed. Firstly, the data set is divided and the training samples are expanded by data enhancement technology. Then, the pre-trained model on ImageNet is transferred to this problem. In order to extract the plate image features efficiently, a new deep feature extraction part is added after the pre-training model, which combines residual connection and depth separable convolution. It can not only enhance feature reuse and alleviate gradient disappearance, but also help to extract the deep features of the image. Finally, in order to improve the robustness of the model in the training process, LeakyReLU function is used instead of ReLU function to avoid neuron death. Compared with other mainstream methods, the model can save the time of zither quality classification and improve the recognition accuracy.

Key words: zither panel, paulownia catheter image recognition, residual connection, depth separable convolution

摘要：

深度学习和迁移学习的兴起为树种识别提供了新方向，然而其在同树种内不同品质间木材识别仍存在挑战。为改善古筝面板品质分级现状，设计了一种深度残差网络模型。首先将数据集进行划分并采用数据增强技术来扩充训练样本，然后将ImageNet上经过预训练的模型迁移到该问题上。为高效提取到板材图像特征，在预训练模型后新增深层特征提取部分，其融合了残差连接和深度可分离卷积，不仅可增强特征重利用率和缓解梯度消失，而且有利于提取到图像深层特征。最后为提升模型在训练过程中的鲁棒性，使用LeakyReLU函数代替ReLU函数避免神经元死亡问题。该方法在泡桐导管图像数据集上测试精度达到了92.8%，对比其他主流方法，该模型可节省古筝品质分级时间，提高识别精度。

关键词: 古筝面板, 泡桐导管图像识别, 残差连接, 深度可分离卷积