焊缝缺陷图像智能分类研究

Research on Intelligent Classification of Weld Defect Image

为解决V型焊缝内部缺陷超声相控阵图像智能分类时代表性特征提取困难的问题,提出了一种改进的稀疏自编码器网络模型(RSAE),获取显著响应V型焊缝缺陷类型的特征集。首先,预处理缺陷超声图像,进而采用纹理特征与形状特征相结合的方法解析缺陷的明暗复杂程度、纹理粗细、沟纹深浅、灰度分布均匀程度;其次,基于Relief-F算法计算各特征对V型焊缝内部缺陷类型的敏感度,分配其为RSAE的初始权重参数,同时给RSAE三种约束实现对样本数据的重新表达。实验使用原始特征与改进的稀疏自编码器编码的特征分别作为核极限学习机的输入,识别准确率分别为87.2%与96.5%。结果表明,提出的改进的稀疏自编码器获得的高级特征较原始特征在模式识别中有更好的结果。

In order to solve the problem of difficult extraction of representative features in the intelligent classification of ultrasonic phased array images for internal defects of V-shaped welds,This paper proposed an improved sparse autoencoder network model(RSAE)to obtain a feature set that significantly responds to the V-shaped weld defect type.First,the defect ultrasound image was preprocessed,and then the method of combining texture features and shape features were used to analyze the complexity of the defect’s light and shade,the thickness of the texture,the depth of the groove,and the uniformity of the grayscale distribution;Second,Based on the Relief-F algorithm,the sensitivity of each feature to the internal defect type of the V-shaped weld was calculated,and it was assigned as the initial weight parameter of the RSAE.At the same time,the three constraints of the RSAE were used to re-express the sample data.The experiment used the original features and the features encoded by the improved sparse autoencoder as the input to the kernel extreme learning machine,and the recognition accuracy is 87.2%and 96.5%respectively.This result shows that the advanced features obtained by the proposed improved sparse autoencoder Compared with the original features,it has better results in pattern recognition.