基于改进多尺度采样算法的焊缝缺陷图像识别

Defect recognition of weld image based on improved multiscale sampling algorithm

为使焊缝缺陷图像特征提取更完整,提出一种多尺度采样分析的焊缝缺陷识别方法。对原有多尺度算法的符号模式进行改进,使其相比原来增加两个划分区间;利用控制变量法确定焊缝图像识别的最优尺度数为2,最佳邻域像素数为12;提取焊缝图像的多尺度特征,连接各个尺度下的特征向量用于表示焊缝图像;用最近邻分类器对缺陷图像进行识别;考虑实际生产中缺陷种类的多样性,对缺陷图像做9个不同角度的旋转。综合实验结果表明,该方法具有较高的识别率且具有旋转不变性,实验中对复杂缺陷图像识别率超过91.71%,优于现有方法,能够满足实际需要。

To make the weld defect image feature extraction more complete,an improved multiscale sampling method was proposed for seam defect recognition.The magnitude pattern(MP)of the original multi-scale algorithm was improved,which made it with more than two division intervals.The optimal values of L and P were 2 and 12 which were determined using the control variable method.The multi-scale feature of the weld image was extracted and the feature vectors at each scale were connected to the weld image.Considering the diversity of defects in actual production,the defect images were rotated at 9 different angles.Comprehensive experiments show that the proposed method has higher recognition rate and rotation invariance.In the experiment,the recognition rate of complex defect images is over 91.71%,which is superior to the existed methods and meets the actual needs.