A high efficiency deep learning method for the x-ray image defect detection of casting parts

Abstract In the manufacturing industry, digital radiography (DR) images are often used to detect internal defects in casting parts. With the development of computer technology, increasingly more researchers use computer algorithms instead of manual inspection. However, traditional computer vision methods are generally not efficient and robust. In this study, we propose a DR image defect detection methodology based on deep learning technology. In order to train and evaluate the deep learning model, we create a casting defect DR image dataset, which includes 18 311 DR images labelled for two types of objects—defects and inclusions. In the methodology, an object detection method baseline named YOLOv3_EfficientNet, which replaces the backbone of YOLOv3_darknet53 with EfficientNet, is used. This operation leads to a significant improvement in the mean average precision value on YOLOv3 and greatly reduces the inference time and storage space. Then, a data enhancement method based on DR image features is used, which can increase the diversity of the clarity and the shapes of defects randomly. To further facilitate the deployment of models on embedded devices with an acceptable accuracy loss range, a depth separable convolution operation is adopted. Regarding the bounding box regression, we perform some relevant research in the training and inference stages of the model, and the accuracy of the model was improved in both stages of them according to the experiments. The experiments proved that every type we adopted could benefit the model’s performance.