基于改进LeNet-5模型的木材表面典型缺陷识别方法研究

Detecting Method of the Wood Surface Defects Based on Modified LeNet-5 Model

针对传统木材缺陷识别方法效率低、精度不高及泛化能力差等问题,对传统LeNet-5模型进行改进:通过分别增加卷积层和池化层的层数至4层,以增加网络深度;采用批量归一化算法,以解决内部协变量位移过拟合的问题;改用Leaky Relu函数作为激活函数,并加入稀疏分类交叉熵作为损失函数,使用Adam作为优化器,来优化网络模型。应用改进LeNet-5模型对辐射松木材常见缺陷(结疤、裂痕)及无缺陷样本集进行识别试验,结果表明:相对于传统LeNet-5模型以及VGG19、AlexNet、ResNet-50三种经典模型,改进LeNet-5模型的训练集准确率最高为99.87%、验证集为99.43%,运算时间缩短,木材缺陷识别精度和效率提高。

In order to improve efficiency,accuracy and commonization,the traditional LeNet-5 model was modified for detecting wood surface defects by increasing the number of convolutional layers and the pooling layers to 4 so that the network depth was increased.The batch normalization algorithm was used to solve internal co-vitiate shift displacement over-fitting problem.Leaky Relu function was used as the activation function,then sparse categorical cross entropy was added as the loss function,as well as Adam was used as the optimizer to enhance the network model.The comparison test was carried out on the samples of Pinus radiata with and without classical defects(knots and cracks).The results showed that the modified LeNet-5 model performed best comparing with VGG19,AlexNet and ResNet-50 classic models.The accuracy rate was 99.87%;the verification set was 99.43%;the operation time was reduced,therefore the accuracy and efficiency of detecting wood surface defects were improved.