【作者】 牟洪波；

【导师】 戚大伟；

【作者基本信息】 东北林业大学 ， 生物物理学， 2006， 硕士

【摘要】 采用X射线作为检测手段,对木材进行无损检测。通过检测透过被检物体后的射线差异,来判断被检测木材是否有缺陷存在。在木材的另一端利用图像增强器进行接收,再经过微光摄像机送入A/D转换器将木材X射线模拟图像转换成一数字图像存入计算机中,运用MATLAB和VC⁺⁺软件的图像处理功能对采集到的木材缺陷图像进行处理和分析,针对木材中不同类型的缺陷,对木材缺陷图像进行特征提取,确定了木材缺陷的尺寸和位置。本文就三种常见的木材缺陷:节子、虫害、腐朽进行了具体的研究。 在无损检测信号处理和特征构造的基础上,运用特征参数建立了缺陷识别的数学模型,针对无损检测信号的特征,构造了人工神经网络,选用多层前馈神经网络模型（BP网络）,网络识别所需要的特征参数能够反映木材缺陷的全部特征。我们把缺陷的灰度均值、缺陷灰度方差、缺陷的长宽比作为进行神经网络识别的特征输入中的三个量,利用反向传播网络的学习算法,对神经网络进行训练。对于某批训练样本,用BP算法,通过反向传播来调整各层神经元的权系数,反复输入所有样本序列,重复以上步骤,直至权系数不再改变,输出误差在规定范围内。网络学习结束后,得到输入层、中间层和输出层各单元的连接系数矩阵。运用MATLAB对已训练成熟的神经网络进行仿真,便可得到输入向量的模型,完成网络识别任务。此过程在完成特征提取的基础上,采用人工神经网络方法对缺陷类型进行有效识别,准确判断木材内部的缺陷信息。实验结果表明这种方法可以成功地对这三种木材缺陷进行无损捡测和分类。此方法也可在对其它木材缺陷的检测和分类上推广使用。更多还原

【Abstract】 X-ray was adopted as a measure method for log nondestructive testing. The difference of X-ray intensity after exposure was tested in order to judge whether the defect of log exist or not. At the other side of log, image enhancement device was used to receive the log image, and then via low-light camera transmit the X-ray log image which was transformed from analog image to digit image by A/D converter to the computer memory. MATLAB and VC++ image processing program were applied to process and analyze the image of log with defects. The characters of image defects were extracted to identify the size and position of defects in a log. In this paper, three common defects which are knot, grub-hole and rot were studied.On the base of signals processing of nondestructive testing and characteristic construction, characteristic parameters were applied to establish the mathematic model of defects recognition, especially for the character of nondestructive testing. ANN(Artificial neural network) was established by selecting multi-BP networks, the characteristic parameter for network recognition could reflect all characters of log defects. Defect gray averaging, defect gray variance and ratio of the defect length and width were served as three parameter inputs for ANN recognition, and the network was trained by using BP network algorithm. BP algorithm was applied to trained samples, and weight-coefficient of neuron was adjusted in different layers by BP algorithm, input all sample sequences repeatedly until all the weight-coefficient no longer change and the error was in the fixed scope. After studying network, coefficient matrix of each unit which includes input layer, intermediate layer and output layer was gained. MATLAB was applied to simulate trained ANN, then the input vector model was gained, and network recognition was completed. Based on character acquisition, artificial neural network was adopted to recognize the kind of log defects effectively and the interior defects information of log was judged correctly. The experimental results show that ANN is an effective method for the nondestructive testing and classifying of three defects. This method can be used in other log defects nondestructive testing and classifying.更多还原