基于迁移学习的木材缺陷近红外识别方法研究

Near infrared wood defects detection based on transfer learning

针对基于近红外光谱的硬木表面缺陷分类检测中,由于样本数量有限、数据代表性不足等原因,导致的分类器泛化能力不足、精度仍有待提高等问题,提出适应成分分析与深度迁移前馈神经网络相结合的方法,有效地迁移其他树种光谱与缺陷对应知识至目标分类器,提高分类器性能。以色木样本为源域,柞木样本为目标域,根据近红外光谱定性分析方法,在900~1 700 nm光谱波长范围内采集了色木、柞木样本表面无瑕、活节、死节等3种近红外光谱数据。经过标准正态变量变换和Savitzky-Golay平滑进行原始光谱去噪预处理后投入建立的基于适应成分分析的深度迁移学习模型。将测试样本投入完成的分类器进行测试,结果表明,当柞木训练集占总样本比例超过30%后,模型对3种类型缺陷的识别率均为100%。

In the classification and detection of hardwood surface defects based on near infrared spectroscopy,the generalization ability of the classifier is insufficient due to the limited number of samples and the lack of data representation,and the accuracy of the classifier still needs to be improved.A method of combining adaptive component analysis with depth migration feedforward neural network was proposed to transfer the spectral and defect knowledge of other tree species to the target classifier effectively and improve the performance of the classifier.Acer mono samples were taken as the source domain and Mongolian oak samples as the target domain.According to the near infrared spectroscopy qualitative analysis method,near infrared spectra of flawless,live and dead joints on the surface of samples were collected in the range of 900~1700 nm.After standard normal variable transformation and Savitzky-Golay smoothing pretreatment,a deep transfer learning model based on adaptive component analysis was established.After pretreatment,the transfer learning model was put into practice.The results show that the recognition rate of three types of defects is 100%when the proportion of Mongolian oak samples training set exceeds 30%.