深度卷积神经网络的X射线焊缝缺陷研究

Research on X-ray weld defects detection by deep CNN

针对X射线焊缝的缺陷识别难度较高且难以分类这一问题,在典型CUDA-CONVNET卷积神经网络(CNN)的基础上,改进并设计了一种深度CNN结构。以图像预处理作为基础,在保证最大限度提取原始图像的焊缝特征的前提下,对CNN的层次架构及参数设定开展了研究;通过与支持向量机(SVM)识别算法对比,进一步评估提出的深度学习方式,研究结果表明:改进后的深度CNN结构及其算法对于大样本的图像特征表达与识别能力有一定的优势,运算样本与错误率成反比,网络结构具有较高的图像分类识别正确率。

Aiming at problem of high difficulty of X-ray welds defects detection and classification , a deep convolutional neural networks （CNN） structure is designed, which is based on improvement of typical CUDACONVNET CNN. The hierarchical structure and parameter setting of CNN are researched under the premise in weld features of original images can be maximum extracted, on the basis of image preprocessed. Meanwhile, further estimate the deep learuing mode by comparing with the recognition algorithm of support vector machine （ SVM ）. The research results show that for large samples, the structure and its algorithm of the improved deep CNN have advantages in images feature expression and recognition ability while its computation samples are inversely proportional to the error rate, which means the larger the samples number is, the lower the recognition error rate is. In addition, the network structure has a high correct rate for classification recognition of images.