基于深度卷积神经网络的木材表面缺陷检测系统设计

Wood Surface Defect Detection System Design Based on Deep Convolutional Neutral Network

针对木材表面的虫眼、结节和裂缝等缺陷,采用传统的人工检测人力成本较高,而采用图像处理技术提取的各种特征取决于人工经验,且受到噪声、光照等外界因素影响较大,其实际应用具有很大的局限性,因此提出基于深度卷积神经网络的木材表面缺陷检测方法。首先构建数据库,考虑到样本有限,采用了数据增强扩充样本数量;其次以在ImageNet数据集上获得优异性能的VGG16模型为基础,采用迁移学习,基于自建的小样本数据库微调该神经网络的后三个全连接层参数;最后利用重新训练好的深度卷积神经网络对测试图像进行检测,结果表明该网络在木材表面缺陷检测上达到了很好的性能。

There are two common methods to detect defects including bug eyes,nodules and fissures on the surface of woods.One is artificial observation by human’s eyes and another is computer vision by image processing technologies.But the former is at high cost of human resources and the latter is confined to specific field because of hand-crafted features and being susceptible to noises and illuminations.A novel technology based on the deep convolutional neutral network model is proposed in this paper.In the first step,the training data set is built by data augment technology to yield medium training data set.In the second step,the transfer learning is adopted to obtain the useful features from certain convolutional levels of VGG16 trained on ImageNet.In the third step,the full connection levels of the deep convolutional neutral network are fine tuned on the self-built targeting training data set.In the last step,the trained deep convolutional neutral network is used to detect the wood defects in the testing data set and the results demonstrate that the proposed network performs well for detecting the defects of woods.