摘要
木材缺陷智能检测技术可以有效降低人工误检带来的经济损失,对提高木材加工智能化水平具有重要意义。提出了一种木材缺陷智能检测算法,通过双循环生成对抗网络(double least generative adversarial networks,DLGAN)及密集卷积网络(Dense-Net)来检测色差、虫眼、裂纹、节子和伤疤等5种木材常见缺陷。首先,使用DLGAN技术扩充数据集,提高数据集的多样性和数量,缓解了因训练数据不足而导致的过拟合问题;其次,基于Dense-Net的特点,采用密集的卷积块序列提高对微弱特征的提取和学习能力,以便更好地检测木材缺陷。试验结果表明,相比VGG16、Inception-v2、ResNet 3种经典卷积神经网络,基于DLGAN增广数据集训练的Dense-Net模型有效提高了木材缺陷检测模型的性能,平均准确率达到92.7%,在只使用少量训练数据的情况下模型依然具有良好的图像生成能力和训练鲁棒性。
Wood defect intelligent detection technology can be used to effectively reduce the economic losses caused by artificial error detection,which is of great significance to improve the level of wood processing intelligence.In recent years,researchers have attempted to apply Dense-Net to pattern recognition by using deep learning techniques.However,the neural network algorithm based on deep learning needs many training samples to achieve high recognition accuracy,and in some cases,the data collection is extremely time-consuming and laborious.It is even more onerous to label these data with information.To reduce the dependence on the data,some studies have used data enhancement techniques to expand the data set.Among them,the generative adversarial network(GAN)has great research potential and application value in the field of data augmentation,because it can simulate any data distribution mode.However,the traditional GAN model suffers from serious degradation when the number of data is unbalanced or the number of data is small,especially when the feature difference is small,the ability of feature extraction is greatly reduced.To solve the above-mentioned problems,an intelligent detection algorithm for wood defects is proposed,in which five common wood defects,namely chromatic aberration,wormhole,crack,knot and scar,were detected by DLGAN(double least generative adversarial networks)and Dense-Net.Firstly,the DLGAN technology is used to expand the data set to improve the diversity and quantity of data set,and alleviate the over-fitting problem caused by insufficient training data.Secondly,based on the characteristics of Dense-Net,the dense convolutional block sequences are used to improve the ability of weak feature extraction and learning,so as to better detect wood defects.The experimental results show that the Dense-Net model based on DLGAN augmented data set can effectively improve the performance of wood defect detection model compared with the classical convolutional neural network of VGG16,Inception-v2 and ResNet
作者
解晨辉
杨博凯
李荣荣
XIE Chenhui;YANG Bokai;LI Rongrong(College of Furnishings and Industrial Design,Nanjing Forestry University,Nanjing 210037,China)
出处
《林业工程学报》
CSCD
北大核心
2023年第4期129-136,共8页
Journal of Forestry Engineering
基金
国家木竹产业技术创新战略联盟科研计划课题(Tiawbi202008)
江苏高校“青蓝工程”项目。
