基于Mask R-CNN的铸件X射线DR图像缺陷检测研究

Research on defect detection of X-ray DR images of casting based on Mask R-CNN

针对传统的铸件缺陷检测不能对缺陷进行分类分级等问题,提出了一种基于Mask R-CNN的铸件X射线DR图像缺陷检测算法。首先对原始图像进行预处理,采用引导滤波进行图像平滑,平滑图像与原图像进行差分得到差分图像,将差分图像与平滑图像相加运算使图像增强,再利用Labelme进行图像标注,形成训练数据集。送入Mask R-CNN深度学习网络,通过特征提取网络生成建议区域,分类、回归网络生成边界框和掩码,经多次参数调节后得到训练网络模型,最后测试数据集。实验数据结果表明,气泡1~5级的检测率分别为:66.7%,71.4%,77.4%,88.9%,87.5%;疏松1~5级检测率为:62.5%,72.2%,77.1%,83.3%,81.1%。检测结果证明应用Mask R-CNN结合引导滤波增强方法的缺陷检测方法可以较好的实现对铸件X射线DR图像的缺陷检测的分级分类,为工业铸件缺陷检测提供了应用深度学习方法的解决方案。

Aiming at problem that traditional casting defect detection algorithms cannot classify and grade defects, this paper proposes a defect detection algorithm of X ray DR image based on Mask R-CNN. Firstly, the original image is preprocessed, guided filtering is adopted to conduct image smoothing, then the smoothed image is subtracted from the original image to obtain a difference image, and the difference image and the smoothed image are added to perform image enhancement. The Labelme is used to mark the image to form a training data set, which is sent to the Mask R-CNN deep learning network and the recommendation area is generated through the feature extraction network. The classification and regression network generate the bounding box and mask. After multiple parameter adjustments, the trained network model is obtained and finally the data set is tested. The experiment data result shows that the detection rates of bubbles level 1 to 5 are 66.7%, 71.4%, 77.4%, 88.9%, 87.5%, respectively;the detection rates of porosity level 1 to 5 are 62.5%, 72.2%, 77.1%, 83.3%, 81.1%, respectively. The detection results prove that the defect detection method using Mask R-CNN combined with guided filtering enhancement method can better achieve the classification and grade of the defect detection of the casting X ray DR image, which provides a solution scheme for the defect detection of industrial castings applying deep learning method.