基于改进深度置信网络的木板表面缺陷检测模型

Improved deep belief network based detection model for wood surface defects

为了提高木板表面缺陷检测精度,采用连续型深度置信网络(DBN)建立木板表面缺陷检测模型。首先,对待检测的木板图片进行关键特征提取,并建立DBN检测模型。然后,将木板图片特征输入DBN的多个受限玻尔兹曼机(RBM)层进行深度训练,从而利用DBN的深度优势来获得木板表面缺陷检测结果。最后,引入人工蜂群(ABC)算法对DBN的权重参数进行优化从而缩短训练时间。实例测试实验结果表明:选择学习速率为0.075时,ABC-DBN算法在划痕、刮痕、裂缝、崩缺4类样本集中的均方根误差(RMSE)均值性能更优。采用卷积神经网络(CNN)、快速区域卷积神经网络(Faster R-CNN)、自适应增强卷积神经网络(AdaBoost-CNN)和ABC-DBN算法分别进行检测准确率对比实验。结果显示,ABC-DBN算法检测准确率RMSE为5.067×10^(-2),是最优结果,Adaboost-CNN算法次之,CNN算法最差。

In order to improve the accuracy of wood board surface defect detection,a continuous deep belief network(DBN)is used to establish a wood board surface defect detection model.Firstly,the key features of the wood board images to be detected are extracted,and a DBN detection model is established.Then,features of wood board pictures are input into several restricted Boltzmann machine(RBM)layers of DBN for depth training,so that the detection results of wood board surface defects can be obtained by using the depth advantage of DBN.Finally,artificial bee colony(ABC)algorithm is introduced to optimize the weight parameters of DBN and shorten the training time.The experimental results show that when the learning rate is 0.075,the ABC-DBN algorithm performs better in the root mean squared error(RMSE)mean value of scratch,scrape mark,crack and flaw sample sets.The convolution neural network(CNN),faster region convolution neural network(Faster R-CNN),adaptive boosting-convolution neural network(AdaBoost-CNN)and ABC-DBN algorithm are used to carry out detection accuracy comparison experiments.The results show that the RMSE of the detection accuracy of ABC-DBN algorithm is 5.067×10^(-2),which is the best result,followed by AdaBoost-CNN algorithm and CNN algorithm.