基于深度学习网络的射线图像缺陷识别方法

Defect recognition for radiographic image based on deep learning network

针对建立射线无损检测智能化信息处理平台的需要,提出一种基于深度学习网络的智能识别方法。以卷积神经网络结构为基础,结合径向基神经网络非线性映射能力,构建一种模拟视觉感知原理的深度学习网络结构,并提出一种网络结构自生长方法和参数学习方法;然后在获取注意区域的基础上,模拟人类大脑深度学习的层次感知系统,使可疑区域的像素灰度信号直接通过深度学习层次网络,通过卷积网络逐层挖掘可疑缺陷区域的本质特征;最后利用径向基网络部分实现对射线图像缺陷的直接智能识别。实验中对复杂射线图像的缺陷识别率超过91%,优于传统方法。实验表明该方法有较高的准确率和较好的适应性,能够满足射线无损检测智能化信息处理平台的需要。

Aiming at the requirement of establishing intelligent information processing platform of radiographic nondestructive detection,an intelligent recognition method based on deep learning network was proposed.Based on the structure of the convolution neural network and combining with the nonlinear mapping ability of the radial basis neural network,a deep learning network structure was built,which could simulate human visual perception principle.And a self-growth method of network structure and a parameter learning strategy are introduced.Then,on the basis of acquiring the attention region,the hierarchy sensing system of deep learning of human brain is simula ted ; the pixel gray signals of the attention region directly pass through the recognition network,which simulates the deep learning hierarchy model of the visual perception system.So the internal characteristic of the suspicious defect region is extracted layer by layer through the convolution network.Finally,the intelligent recognition of the defects in the radiographic images is achieved through the radial basis network.The experiments show that the recognition rate of the defects in complex radiographic image is above 91%,which is superior to that of traditional methods.So this method has higher accuracy and better adaptability,which can meet the requirements of the intelligent information processing platform of radiographic nondestructive detection.