基于优化卷积神经网络的木材缺陷检测

针对深度学习中的卷积神经网络算法,在木材无损检测过程中存在缺陷定位不准确、缺陷轮廓和边界信息不完整、识别精度需进一步提高等问题,利用非下采样剪切波变换最优稀疏表示特性,以及简单线性迭代聚类算法能很好地保持像素紧凑度和图像边界轮廓的优点,设计了一种优化的卷积神经网络算法,以提高木材无损检测的准确率。首先采用非下采样剪切波变换对采集的木材图像进行简单预处理,保留木材图像的缺陷特征不丢失,降低图像处理的复杂度以及运算量;然后利用卷积神经网络对木材图像实现深层次的算法设计,同时应用简单线性迭代聚类算法对初步模型进行增强改进,提取出相对准确的木材缺陷轮廓;最后通过反复调整参数和调试优化器,优化卷积神经网络算法的收敛速度,提高学习和运算效率,完善卷积神经网络对木材缺陷轮廓的提取,在降低运算复杂度的同时,提高其精度,具有良好的鲁棒性。相比径向基函数(RBF)神经网络、向后反馈-径向基函数(BP-RBF)混合神经网络和卷积神经网络,本算法对木材缺陷具有更好的识别效果,其识别准确率达到98.6%左右,且识别时间相对更短。

Wood defect recognition based on optimized convolution neural network algorithm

Deep learning technology is a hot spot in machine learning research at present.Establishing and simulating the neural network of human brain to analyze the characteristics of data information were extracted to imitate the working way of human brain,which has shown great advantages in image processing.In this paper,with the help of the optimal sparse representation characteristics of non-sampling shear wave transform and simple linear iterative clustering algorithm,the pixel compactness and the edge contour of the image could be well preserved.Then an optimized convolution neural network algorithm was developed to improve the accuracy rate of wood nondestructive testing to solve defect localization inaccuracy as well as contour and boundary information incompleteness,which further improved the defect characteristics recognition accuracy.Firstly,the non-sampling shear wave transform was used to pretreat the collected wood images,which can reduce the complexity of image processing and the amount of computation while retaining the defect features of the wood images.Secondly,the convolution neural network was used to design a deep algorithm structure for the wood images.At the same time,the simple linear iterative clustering algorithm was also used to improve the initial model,from which the relatively accurate wood defect contour was extracted.Finally,the convolution neural network algorithm was optimized by adjusting the parameters and debugging optimizer repeatedly to improve the learning and computing efficiency,and refine the extraction of the wood defect contour.The optimization improved the processing precision with the reduction of computational complexity,and it had a better robustness.In addition,this algorithm has an excellent recognition effect on wood defects,and the recognition accuracy reached 98.6% while the recognition time was relatively shorter compared with those of the radial basis function (RBF) neural network,back propagation-radial basis function (BP-RBF) hybrid neural network and normal convolution neural network.