基于X射线图像和卷积神经网络的石油钢管焊缝缺陷检测与识别

Detection and identification of SAWH pipe weld defects based on X-ray image and CNN

研究了基于X射线图像和卷积神经网络（CNN）的石油钢管焊缝缺陷检测与识别问题。首先采用数字图像处理技术拟合提取出焊缝区域,验证了咬边缺陷的存在不影响焊缝边缘的提取;针对常用阈值分割方法不适于小面积区域缺陷分割的局限,采用基于排序点的聚类算法（OPTICS）对区域内任意形状大小的缺陷和噪声干扰点进行分割,然后对缺陷、噪声和无缺陷的正常图像进行提取并进行数据增强及尺寸归一化操作,从而完成焊缝图像的预处理以构建出样本图库。最后采用CNN与Softmax分类器相结合的算法,以缺陷和噪声为输入样本训练CNN并进行了实际应用实验,实验结果验证了方法的有效性。

This paper focuses on the detection and recognition of the weld seam defects in the Helical Submerged Arc Welding（ SAWH）pipe based on X-ray image and convolutional neural network（ CNN）. Firstly,the digital image processing technique is utilized to detect and fit out the weld seam area of SAWH pipe X-ray images. The detection of weld edge is not affected by the presence of undercut defects. The common threshold segmentation technique does not fit the segmentation of small-area defects. To deal with this limitation,the ordering points to identify the clustering structure（ OPTICS） algorithm is adopted to segment defects and noise points of any size in the weld seam region. Then,the defect,noise and normal images are extracted. Data enhancement and size normalization are performed to complete the pretreatment of these images. The sample image database is built accordingly. Finally,the recognition model of welding defects is trained by the combination of CNN and softmax classifier. The defect and noise images of practical application are utilized as input. The effectiveness of proposed method is verified.